Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

The AI Native Dev - from Copilot today to AI Native Software Development tomorrow

The AI Native Dev - from Copilot today to AI Native Software Development tomorrow artwork

The Background Check You Can't Run on an AI Agent

Aug 11, 202648 minAudio
ShareXLinkedInFacebookWhatsAppEmail

The thing that makes an agent useful is the exact thing that makes it dangerous. Keycard co-founder Ian Livingstone breaks down why non-determinism is both the feature and the bug, and why identity, not model quality, is what really gates how much autonomy you can hand an agent. If you have ever clicked "always allow" without reading it, this one is about you. What we cover: - Why authentication was enough in the cloud era and stops being enough with agents - The background check you can't run on an agent, and what has to replace it - Mission identity: who is acting, on whose behalf, and for what purpose - Cross App Access, Agent Auth and the protocols trying to fix OAuth - Consent fatigue, LLM as a judge, and where hard boundaries still belong - Why MCP ships with an auth story and CLI tools don't Chapters: 00:00:00 - Introduction 00:03:43 - Why every wave of computing rewrites identity 00:04:50 - The feature and the bug are the same thing 00:10:23 - Why shared secrets break for agents 00:13:36 - The background check you can't run on an agent 00:19:42 - Chargebacks, delegation and proving intent 00:22:16 - Mission identity: a new layer 00:28:07 - Cross App Access, Agent Auth and emerging protocols 00:33:02 - MCP vs CLI, and how Keycard works 00:43:22 - Identity three years from now 🌐 Tessl: https://tessl.io 🔔 Subscribe for weekly episodes on AI-native development Where do you draw the hard line for your own agents? Tell us in the comments.

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

SPEAKER_02Ultimately, the identity question is, how do I ensure when an agent does make a mistake, it doesn't do something catastrophic I didn't intend? This thing is built on like a probistic distribution. The feature is that it can reason and it can guess over large amounts of data, right? Um, and that it is non-deterministic, that's a feature.

SPEAKER_04Right.

SPEAKER_02The bug from a security perspective is that it's non-deterministic. Yeah, because the feature and the bug are the same. And the final question is how do I know that an agent is performing actions aligned with the intent of whoever set the agent to do something?

SPEAKER_00The AI Native Dev is a podcast with developers and engineering leads at the cutting edge of AI and agentic coding. Join your host, Glyfer Johnny, and me, Simon Mayfor, every week as we chat with the most exciting voices in AI and tackle the biggest questions facing developers today. This is the AI Native Dev. We just wrapped up two amazing days at AI DevCon in London.

SPEAKER_01But the great thing is that we get to do it all over again in New York City 5th November.

SPEAKER_00You're absolutely right. We're gonna be back in a city that never sleeps on November 3rd and 4th for more amazing sessions, really engaging, hands-on workshops, and much more.

SPEAKER_01Yep, all that great networking, partying, eating and drinking that you've come to expect from AIDEF CON. We think we have one of the best hallway tracks in the business, and it's the perfect compliment to our incredible speakers and presenters.

SPEAKER_00We'll both be in person and virtual with live-streamed access to all mainstage keynotes and talks.

SPEAKER_01Sign up right now for our Super Blind Bird ticket for just$100, only available for a limited time.

SPEAKER_00We're really excited to be headed back to the big app. We hope to see you all there.

SPEAKER_03Hello everyone, welcome back to the AI Native Dev. We are in uh kind of an ad hoc office over here, you know, actually sort of in uh hometown or like a home base, home base at the moment, uh here in San Francisco as part of the AI Engineer event, uh, and really you know, kind of enjoying a lot of AI conversations on it. And today we're gonna dive into identity. You know, identity if you've been sort of exploring the security realm of agentic uh anything really uh has been you know the top uh topic of it.

SPEAKER_03You know, everybody appreciates that it is a foundational aspect of how do you get these agents to actually be secure and kind of work and give them some autonomy. And so to kind of demystify this whole domain and understand how to tackle that, we've got Ian Livingstone, who is the co-founder of Key Card. Yeah. Uh uh that is uh a sort of a really exciting kind of company in the identity space. Uh, but also really Ian is really sharp and deep around explaining things as a whole, and specifically identity. Uh, I've known him from uh sneak days and his sort of past uh companies that is founded on it.

SPEAKER_03So we'll uh we'll learn about identity, we'll learn about what you can do, and we uh we'll hear a little bit about Keycard in the process of it as well.

SPEAKER_02Yeah, thanks for coming on. Thanks for having me. I'm super excited to talk about identity and agent and security and this whole bundle of things together today, and uh really help people understand sort of how one is like what's the problem, um, why is this a hot topic, like why is identity, like identity is like kind of boring, it's snooze fest, it's been for years, you know. But why is this hot again and why are people thinking about it and how that actually changes the way that we actually think about the way we build systems and also how autonomous these systems can become. Yeah, yeah. So well, let's unpack over here.

SPEAKER_03So let's let's get started. Just tell us a little bit about what what are the uh the problems or sort of the you know, give us a bit of a structure for thinking about identity and agents.

SPEAKER_02Absolutely. So I think I think the first step we have to think about is like why is identity even a thing we talk about? Like if we were to step back, and like identity has been a part of computing for as long as it existed, right? The minute you had a mainframe and then I wanted to have multiple users, now you have an identity problem, which is like what files or programs on the system are guy's programs and files, and what files and systems are Ian's, and how does Guy have access to all of Ian stuff, or does Ian have access to all of Guy? Like it's been like every wave of computing introduces this question of well, who can do what, when, and where? And who can access.

SPEAKER_03Exactly. I remember like uh high school sort of mini hacking, getting into all sorts of information systems that I was not supposed to be about.

SPEAKER_02Exactly. And you know, in the rise of the internet brought about, like when we uh you know, initially it was like TLS brought about commerce, and so every wave of computing is comes with that net new security challenges that we have to solve because it changed, like the interaction patterns change when we introduce like new paradigms of the rise of the cloud, and then etc. You know, with IAM and identity for for cloud vendors like an AWS and GCP and secrets for all, we've always had these problems. And so agents, like many things, as we introduce new functionality and capability like platform shift, we have to reassess our security model.

SPEAKER_02And I and you know, the core crux of the identity problem when it comes to agents is this thing is built on like a probabilistic distribution. It's a statistical training, right? It's beautiful. The the feature is that it can reason and it can guess over large amounts of data, right? Um, and that it is non-deterministic, that's the feature. Right. The bug from a security perspective is that it's non-deterministic and that it's still a feature. So the feature and the bug are the same. And so the fundamental question is how do I know that an agent is performing actions aligned with the intent of whoever set that agent to do something, right?

SPEAKER_02Which um is like the fundamental question of sick of this generation, is it and there's many different hacks of that, right? You got supply chain security is a part of it, you know, data security is a part of it, uh, but identity is very foundational in thinking through how do I have a secure agent. And it's also foundational to think through how do I enable an agent to be autonomous, right? Um, and the more autonomous an agent becomes, or the more autonomous you want an agent to be, the security equation becomes much more difficult to solve.

SPEAKER_02And the systems that used to work in last last era start stop working as much in the new era, right? And that's I think that's a fundamental sort of like thing to understand. And then ultimately the identity question is how do I ensure that when an agent does make a mistake, it doesn't do something catastrophic, I didn't intend, right? And so today, if you've you know have paid attention or have read, there's lots of different agentic problems associated with identity that people talk about. One of them would be you have supply chain attacks that result in database passwords and keys being leaked, like the most recent Light LM and attacks, that the actual attack factor was how do I take long-lived API keys off of disk and use those to exfiltrate data.

SPEAKER_02But there's other problems of um how do I, you know, when I use an agent, if I ask an agent to do something, this is very common six months ago, was I asked the agent to optimize a database, a potential path to optimizing a database, it would actually be overright. Exactly, right? And so it's snappy. It makes it snappy, but that would be deeply unaligned with what the actual user wants that agent to do. And so that's finally a question of um of a Danian access. Yeah. And the difference between what we've had and what with agents is I actually want to give agents access to different things depending upon the job I've assigned them to do or the task I've assigned them to do.

SPEAKER_02And that's a very different model than we've had previously. Yeah.

SPEAKER_03So I think there's there's basically a few aspects that you that you outlined over there. So one aspect of it is is um is the the non-determinism, right? Or the safeties, like how likely is the agent to uh sort of choose to do a thing that is not a thing that you would have otherwise kind of reasoned and is allowed on it. So I guess there's sort of a broader safety element, and some of that comes down to you know harnesses and context and all that of trying to sort of steer the agent to understand the lines are. And I guess that's that's not really identity, those are like agentic workflows.

SPEAKER_03But then next to that, I think we have sort of two aspects of identity that you touch on there. One is uh, you know, who is the agent, you know, what is what is the sort of entity, right, that I'm signing it, and and the second is the the uh permissions, you know. I guess we talk about authentication and authorization in identity. Um so I I guess how do you am I kind of dividing that correctly uh over here? And maybe if we put safety is a massively interesting topic on its own.

SPEAKER_03Yeah, but do you think they're intertwined or the and or do you just need to worry about all of them?

SPEAKER_02I think I think it when if when a business, when you step back and ask, is it safe? Yeah, or another way to think is do I trust an agent be able to have perform this action? Yeah, it you know, identity and access, uh delegation, uh all these things are fundamental components of that, as they are in every other era of computing.

SPEAKER_04Yeah.

SPEAKER_02And you know, to a certain extent, you when it comes to identity and access, there you'll have deterministic controls around what agents can do, and you'll have non-deterministic controls. And like a good example is like safety work, often is about how do I bias a model into ensuring that it is it is not affected or has you know does the outcome I want. So it's not gonna do the bad thing or has less probability to do the bad thing. But from a security equation, you want to say, well, there's a set of things I just want to never happen. I want to guard, like I don't want a guard rule, I want to I want a hard boundary around what it can do.

SPEAKER_02And it's part of this is also the reason why things like sandbox is very popular right now. It's like I can never 100% know the model may not do this. Right. So, but I need in order for me to put this in production for this type of workload or perform this type of task, I need to ensure that it has a hard boundary so it can never happen. So I can say as a business that yes, you know, the worst case scenarios that I worry about have actually been removed from the equation, and then the rest of the things that are left, it's like kind of okay if that happens. Like it's not great if it does something weird, but it's not gonna destroy the business or result in lost customer data or on and on and on it goes.

SPEAKER_03So you there's uh one aspect of it is defining the sandbox. So you're saying I know that these things won't happen. Exactly. And you know, if I want the agent to do useful things, like if it might not delete all of my inbox, but I'm still getting it to send emails, exactly. But it can still do unsafe things over there. Yeah, but if it doesn't have access to my AI records, it's not gonna send those AI records uh outside. Okay, so so there's we're we're talking now a little bit more about that sort of boundary zone. Exactly. So what are the types of agent uh identity conversation that we have?

SPEAKER_02Yeah, I think I think there's one is like sort of where we're coming from, right? And if you look at sort of the identity systems of last generation, like the way I would give a piece of software access to something is I would give it an API key. It would maybe, you know, service account would represent a service account potentially, it would maybe represent a user because it was a user's API key or personal access token. It was like we built a world based on shared secrets. Right. And and the issue with shared secrets is they often bound strictly to a specific identity, and they bound specifically strictly to a set of permissions associated with that long-lived secret.

SPEAKER_02And the more important thing is if that secret were leaked, anyone who got access to that shared password also could do that thing. And now the the fundamental challenge is okay, knowing that that's where we came from, how do we move to a world where I want, you know, I'm guide the journey, I want really want this agent to operate in a very long period of time on my behalf, because it's gonna have access to my Gmail or I have access to my Google Drive. Um, but if it at any point it needs to perform a right action, so send an email on my behalf, I should probably approve it, right?

SPEAKER_02And how like fundamentally the process we're going through is as we go up the autonomy curve, so we go from I love to use like self-driving car analogy, as we go from level zero, which is you know to terms six off for pre-chat GPT, to you know, co-pilots, which is like you know, level one, yeah, to something that's like slightly or tab complete from cursor, to sort of like cursor agents and claude code agents, which are like level two, is a human still involved. How do we actually go from a world where like as we go up that autonomy ladder, I'm retaining security over what that thing's doing?

SPEAKER_02And if you know, currently, if you were to use cursor or claude, when it goes into a tool call, oftentimes they're saying, Hey guy, sure you want to do that? Yeah, you'll always want to do that? Yeah, no, you don't want to do it in on everything. And so the question is you can't have autonomy when the human has to constantly be in the loop, which is where these access systems come in, which is saying, as I go up the autonomy ladder, so as we want to get to this sort of self-driving car analogy, there's a trust and safety equation then associated with that, which is how do I keep this thing driving on the road instead of like driving off the road off the cliff or whatever has to be done.

SPEAKER_02Yeah, and that's at the crux of it what we're actually all talking about. Right.

SPEAKER_03Yeah. So I I feel like uh uh the word identity is like uh can be a little bit confusing here, right? Because there's like one question is I've got this agent, it's running on my behalf. Uh can I know that it is the agent? Yeah. So like I feel when I hear identity, I I think primarily about who is it that is acting. And that might be uh, you know, is it me talking or is it my agents talking? So a system can separate the two of us. It might be, is my agent talking to another agent to perform actions on its behalf? So it's kind of you know, agents all the way down, right? They uh they cascade down.

SPEAKER_03So that's one aspect. But a lot of the attention you draw is actually to uh to the uh kind of to its decisions almost, right? Like you know, as it as it comes along, um it might be it's around the communication of when I delegate a task, what is not just the uh the knowledge that it is my agent, but rather what is its scoped permissions right now within this world. Like you can compose the email, but you are not allowed to send the email. Exactly. So is that I guess people like again, the word identity is used for the market on it.

SPEAKER_03Exactly. How do you separate this sort of authentication from authorization? Yeah.

SPEAKER_02So this is interesting, right? Because if you look at in the identity world, the cloud when we getting to the cloud was really about can I identify, so authenticate that guy Pajurney is in fact guy Pajerney when Guy goes to Google Drive? Like that, the basis of cloud is really focused on that, and authorization didn't really fall into the problem space. And that's because at the time we didn't really have to solve authorization. And we didn't have to solve authorization because if I could identify that it's Guy and I trust Guy, um, then I can give Guy really broad base access because implicitly we've he's we've done the background check, right?

SPEAKER_02We called all the references, yeah. He you know isn't uh isn't a sociopath, right? Like it turns out well. It turns out that Guy really you know cares about how people think about him and he wants to do a good job, and it turns out if he got fired from the job, that'd be devastating. So that implicitly we can we can trust that guy is going to operate with high intent and not be malicious, right?

SPEAKER_03And part is because it's the same guy that will sort of come into the next task and then S that things in the next task while the agents are basically a brand new creature every time.

SPEAKER_02Exactly, a brand new creature every time, a completely different context, and they're they don't have um they have no sense of is this good or wrong. Like they're not malicious, yeah. Right? They they just don't know, and there isn't a way for them to say it, do I know or do I not know, right? Um that's authoritative in a way that a human can be. And so from that perspective, agents really change that equation where it used to be that, okay, if I just know it's a human and I trust that human, once it's all configured, it kind of works. Now it's based on the thing that I've assigned them and the data they have about that thing, I'm going to give them different levels of access based on what's in that context window.

SPEAKER_02And so much about this next generation is it is about authorization, which is how what things can this do, but it's also about what things has this agent seen or done. So I know that it is in a well-bounded box on whether it should have access to that. And if I can't call that judge, if I can't make a decision, like a system can't make the decision, system can also be like, hey, you know what? Actually, maybe I need to go talk to Guy and be like, hey guy, can you review this thing that the agent wants to do before it goes and does it? And so much of if we're if what we're trying to do is to basically turn software into the Waymo experience, so much of it is like, how do we, as the models improve, how do we over time give them more autonomy, but with the same substrate system?

SPEAKER_02Because there's gonna be certain tasks that you assign an agent that you're gonna say, you know what, it's totally summarization where I'm pulling in a bunch of context, I'm summarizing a document for you. You should have, it's okay, read access to everything. That's totally fine. But the minute you want to go and do a high, maybe a transaction with a cost of over$500, definitely that's a maybe that's when a human has to say, you know what, I don't trust agents to make decisions on my behalf over$500, or organization may say, I don't trust agents to have access to customer data at all. Like there are different trust equations. And every person and organization, depending upon the context of what the agent is, who's using it, and what things it's re-asking, will have a different equate trust equation about what they think is okay for them.

SPEAKER_03Yeah, yeah. Interesting. So really what you want is you want the uh the identity, or really like this identity plus authorization bundle to be task-based, to be I'm performing a task, this task should have a set of permissions. Exactly. Um I think within that world, it feels there are two uh two simplifying a little bit, two ways in which sort of agents run. One is uh an agent as in an agentic workflow. And so I'm running it, and so you know, I can choose. Sometimes it's doing broad things, sometimes it's doing narrow things.

SPEAKER_03Yes. But I think within that world, it feels quite kind of manageable to say I'm giving it authorization. We'll come back a little bit to how do I define which permissions it should have. You know, that is sort of a hard problem. But but at least I know like this bundle, this sort of this sort of this process here that is executing, it's it's kind of executing the same thing or like this assigned action again, again. Um, and then I've got like you know, my kind of Claude Code or Codex or Gemini that I'm sort of running locally.

SPEAKER_04Yes.

SPEAKER_03Uh and I don't know, like probably like in my in my reality or like claud that I'm running on my desktop, I I don't, it doesn't do one thing. Like I do one thing and then I come back, it's just like I've got this other one, and I oftentimes just sort of run it in like one long session that I walk around. And you know, if I just sort of approve a thing, it's like okay, I just performed an action and says, Can I delete this file and delete this file? And then it's like fine, okay, delete all, like yes, always approve, and then I come along and I do something else, and you know, that goes away.

SPEAKER_03Is that a lost cause? Like, is that uh it's like, hey guy, you just stop using uh the agent like that, like use it in a more methodical fashion and delete that, or or or is there a better model for HUD to head?

SPEAKER_02There's definitely a better model. Like what we want to get to is a world where agents, you can assign agent a task, a task maybe something, hey, do like survey the internet and every day come back to me, or survey all of my assets, all of my inbox, and come back to me and tell me what happened, right? And that's like that's a long-lived process where you definitely don't want to every morning wake up and have to click a consent screen, right? You you actually just like I've assigned you, I've given, I've delegated authority to you to perform things on my behalf and work in order for me to delegate to this agent, I and that agent actually, and this is where identity comes in.

SPEAKER_02We have to be able to identify the agent, right? So it's like I've given FUBAR agent for the contact for this task, this authority. And when I sign a different task, I wanted to give it different authority, but that's the first basis. It's like this is why we talk about identity is okay, well, in order to give something your authority to do something on your behalf, you both have to be able to be able to identify that thing. And then the thing that the downstream systems that are saying, hey, is this random request that's coming into my system? Does it have the permissions? Like, has it been given the authority to do the thing it's asking to do?

SPEAKER_02Well, I have to both know, well, oh, Guy did in fact give FUBAR access to do this thing, but I have to actually be able to identify both FUBAR and Guy. And certainly those systems want to treat an access request or a request based from Guy differently than they do from FUBAR. Because when you think of like, I love to give the example of um like credit card chargebacks here, because if you ask your agent to perform uh uh a transaction on your behalf, and then you say, hold off a second, I didn't give it access to do that.

SPEAKER_04Yeah.

SPEAKER_02How does the intermediary be like, well, actually, no, we have a record uh, in fact, that you did tell FUBAR to go and do this thing, and then you said it could spend up to$500, and it did do that thing, so we're actually not gonna process that chargeback because you did do it with today's identity systems, with what we built in the last generation, you can't differentiate between, well, did Guy give FUBAR that access for this specific task or a different task, or where's FUBAR in the mix? None of that, none of that existed because we actually didn't have to build that for last generation security posture, and now we we do.

SPEAKER_03Yeah. So uh, I guess kind of echoing back a little bit, and we'll get a little bit to solutions. Uh I'm hearing three levels here. You know, one is uh separate kind of me from my agent and so use identities that are specific for that agent. Two is uh have a set of authorization that is um that is aligned to the task that you've just given uh the agent. Uh and I guess three, which relates a little bit to the two, is if you're within like the the second one is easy when you've defined the task up front and it's not a long-lived agent.

SPEAKER_03But when it's a long-lived agent, uh am I right in understanding that uh you're saying there's another kind of bucket of like ephemeral entity, the sort of entity that comes along, which is what is the current task being done? Exactly. Uh, that is like kind of like gathered out of the moment in time, the last few messages, whatever it is, you know, the last message that the user has provided, whatever it is. Um, and I guess that type of activity, uh, you know, it really isn't a place in which the human will engage.

SPEAKER_03So that probably needs to itself be more kind of inferred or decided. Yeah. Is that right? So the first one and the second one are a scale problem of a thing that we had before. We had a system, we gave the system a task, we had to give it an identity, we had to give it uh uh a permission set. But this third one, that's a new creature.

SPEAKER_02Totally. And and and if you were to go it read. Current literature, you'll find that what you and I are talking about, like I think tasks is very understandable for us to think about is like, oh yeah, I assigned this thing a job or a task to go do, and I went and did it on behalf, and maybe that task is repeats every 30 days, or it takes a long time to do that task. Um, in literature, it's called a mission. Um, is like often in standards that are starting to emerge, there's that concept of a mission. And the mission is over time, as users or other people give assigned work for these things to do, you're going to think about here are the missions or the quests I've assigned to this thing.

SPEAKER_02And I as a user will update what that mission is as we learn things. Right. And really, what that mission is is about describing the user's intent as a concrete thing. It's like I intend for this agent to be able to, you know, complete this type of task, and that allows another system, like an authorization system, to understand hey, this user or this company, right, somebody has assigned a mission to this agent. That mission is something I can measure their requ access requests against and say, is this aligned with human intent?

SPEAKER_02So that gives us a way to like think about how the concept of uh an agent's identity and their and the task they're working on travels across systems, right? Which is the issue that doesn't exist at all today is if you go to ABOS and you ask your agent to talk to ABOS on your behalf, ABUS has actually no way to know whether you what the action that the agent is performing was aligned with what your initial request was at all.

SPEAKER_04Right.

SPEAKER_02And so part of this is like we have this new concept as in delegation is this what is it that this thing has been assigned to do so that I can then judge whether the thing it's asking to do is aligned with the thing it was assigned to do. Right.

SPEAKER_03So this is like a so we have human identities, we have non-human identities, and is there now a mission identity? There there is.

SPEAKER_02And it's uh ta typically this is a s there's there's in it in you can think of these concepts of inside identity, there's been this concept of session for a long time, right? It's like I log into a website, that creates a session, often from you know 1990 to 2000 year web, it's a cookie in your browser, yeah, and that represents some session, and that session helps identify who you are. And now it's not just a session, isn't just um who you are, it's who are you acting on behalf of, yeah, and it is also for what purpose.

SPEAKER_02Yeah, right, and that's the the three layers that we now have to think about, and for that purpose gives you access to do what things. Yes, yeah. And that's those are the layers of complexity. And if you if you think of it from that perspective, that's how the layer of the complexity of identity systems have are that that agents force for us to be able to solve a lot of these safety concerns across the system.

SPEAKER_03Yeah, okay, cool. I love the session analogy because it's something you can kind of relate to on it. And sessions are interesting because sessions are things that they might get invalidated and require a new one for a variety of triggers. Yeah. Most common time, you know, like it's been idle for a while or maybe just you know, sort of sitting there. Sometimes it's uh an action that has been performed that kind of requires a reauthentication and a variety of those. So maybe let's sort of go into indeed a little bit of the solution landscape on it. So, what are what are the types of sort of tools or new capabilities that are out there? You know, I know Kickard is kind of one of those on it, but give us a little bit of the families of uh ways to tackle this sort of new challenge.

SPEAKER_02Yeah, I think there's there are many different sort of solutions on on market that kind of message to this type of thing. But I think like let's step back and think about what are the new technologies or protocols that are emerging to start solving this, and then we can kind of discuss about like different ways. So I think there's there's really if I'm sitting in the chair as a security person, I'm thinking about two things. One, I'm thinking is about how do I find my shadow agents or my shadow entity? It's always a conversation. Security is what are the things that are happening that I don't know about, right?

SPEAKER_03So I can quantify people are obsessed with shadows. Shadows is like, yeah, can I see a shadow?

SPEAKER_02Is a user doing something I didn't intend, or is a system misconfigured, or whatever, right? This is this is governance, and there's a lot of money made in governance and security, which is quantifying effectively risk. And then there's the golden path, right? And in all security solutions, in all dynamics, you always have, hey, here is how we're gonna find all the risk. Well, once we've identified risk, because a lot of the risk is not people doing the wrong thing, like they don't have the wrong intention, they just don't know how to do the right thing that is like the secure pathway. And then you have your solutions that are what are the secure pathways, yeah.

SPEAKER_02Exactly. And so you always in most markets and in most problem spaces, you have well, let's go find and quantify all the risk, and then you have, okay, here's a place we we go. And over time, as markets consolidate, they often tend to be the same thing. But move them from the shadow to the road. Exactly. But and and the road is typically the infrastructure part, right? And we saw that with the rise of social on the web, you know, with the rise of OAuth, you things like OSero come along and made it really easy for developers to build um social login on their website and user management, which turned out to be a really hard problem. And so a lot of people, that's why that market exists.

SPEAKER_02And then, of course, we had for internal workflow, we had like things like Okta that rose, and it made it real easy to have a golden pathway for me to log into stuff. Right. Um, and then you, of course, had all your governance functionality that came as well. Right.

SPEAKER_03These are all kind of examples of places where you would sort of centrally define your sort of identity, your authentication, like your system, but then you'd be able to kind of plug that in to many places.

SPEAKER_02And and even with the rise of cloud, we had things like Hash and Corp Vault and Terraform that allowed me to configure IAM systems or put a place for me to store long-lived secrets. And so much of the last error was about where do I, where's the golden path. Right. Um now with the change to with the rise of agents, agents basically break a lot of assumptions of how all those things work together because they kind of work across different systems. They can work on behalf of your customers, they can work on behalf of yourself, they can work on, they can just be pieces of code that are running in your stack, you know, maintaining your software factory, deploying, debugging things.

SPEAKER_02You have all these different interactions that are no longer human-driven, but they do exist, and how do you manage that? Um, but the fundamental issue is what are the new technologies or techniques that are coming out to help do the session mission per tasking? And so in OAuth, you have some movement there that's trying to trying to take over OAuth. Some of the stuff that Oxford just came out was called cross-app access. It solves some of the problem equation, but doesn't solve all things. So that's ID Jag for those who are watching.

SPEAKER_02And then on um, there's a net new protocol called Agent Auth, written by Dick Hart, who um is a net new approach on a protocol to try and solve these quad these mission, session, orientation, and loop and move along with secrets. So those are some of the protocols that are emerging. And then the different approaches uh today are you're gonna is how do I find and quantify what risk I have, whether it's on my endpoint or my production stack or my cloud accounts, and then you sort of have your goal path things of how do I actually enable my developers or my customers or my employees to be able to find, build, and use agents uh successfully in sort of the golden path.

SPEAKER_02Yeah.

SPEAKER_03Okay. So let me kind of uh echo back some a little bit of that. So like the OAuth, so you have the sort of broader challenges. One of the projects around OAuth, it's about sort of allowing agents, identifying as agents, not sort of masquerading as humans to use OAuth. Today, you know, I guess we will have all seen, hey, it ran the login, it sort of opened up your browser, you're kind of clicking a thing on it. This is you authenticating as a human, then giving that key now to the agent. Exactly. Instead, can we sort of create a protocol that allow the agent allows the agents to sort of authenticate as it as an agent and sort of have the appropriate?

SPEAKER_03Okay, so that makes sense. And that's like a a bit more of a uh uh sort of like a technical path with sort of the correct identity, but it doesn't solve the sort of the the uh the mission sort of identity like problem that we talked about before. I guess the sort of the second bucket is more around that sort of mission identity. Exactly. And for the mission identity, like where is the point in time in which and and and like who is involved in saying, fine, there's a mission, you've done it, but like what are the permissions that are allowed over here?

SPEAKER_03It seems untenable that that would be a human uh that will come along every time and say, okay, fine, I figured out that what you're trying to do is compose an email, you'll have these permissions versus you know you're optimizing a database, so maybe you are allowed to modify an index.

SPEAKER_02Exactly. I think there's there's two parts to this. One is everybody, individuals and companies, more companies and individuals, are going to have things that you say, you know, no agent can do this without me being involved, right? For me personally, it's probably spending over a certain amount of money or deleting data, right? Like those are like if you're deleting, I want to I want to make that decision, right? Because I care. Or if you're sending an email, I I actually do care deeply.

SPEAKER_03And those are easy, because those are like not mission decisions, those are like just ground wide decisions.

SPEAKER_02Yeah, they're just ground rules. Now, when it comes to mission decisions, this is really where fit patterns like Elm as a judge start to start to emerge. Because you're you're using you know a reasoning engine to reason about whether something's reasonable, and then ostensibly you basically based on does the system think this is reasonable? Okay, if it's not reasonable, now who decides what we do as a result? Is that no, absolutely not, because it's hit some determinants of guard rule. We say no, we'd never let delete. Oh, it's trying to do X, right? Should this go to the security team? The security team then has like a system that reviews it, or should it actually go back to you know the person that's operating on behalf of and says, hey, does Agent try to do this thing?

SPEAKER_02Yeah, is this aligned with your what you were trying to get done? Right? And the biggest challenge we have today when it comes to age genetic security is how do we actually not end up with consent fatigue, right? Like the worst part about the yes, no allow always product dialogue is you just click yes because I'm not gonna read it. Yep, right? And that's always been one of the biggest challenges in security is how do I have a security system that is only sounds the alarm when a program is relevant. Exactly. Otherwise you lose all the signal. So the mission componentry, one part of it is getting all the protocols and the functionality in place so that you can sort of federate out what is it that this person said this thing's actually supposed to be doing.

SPEAKER_02Um, and then the second component is the system around that that can actually make judgment calls. Yeah. And the better the system around it is at making judgment calls, um the more like the less interrupts the users will have, yeah, right, and a lower risk profile decision we made, and thus more trust and a pathway to higher autonomy.

SPEAKER_03Yeah, yeah. Makes sense. Okay, so you have to do that. So again, agents all the way down, you know, something into it. So you need though the infrastructure, we have these two open protocols that I guess are forming, right? Like standards take a moment. Yes. Uh, and uh and then you have uh uh I guess kind of this sort of new concept, which is you have these missions, you have to give them an identity, which is already like a little bit hard to sort of uh identify where that is, but then also you need to figure out what is the authorization to give them, what are the permissions, which will be a subset of your overall agent identity permissions.

SPEAKER_03So what are the the let's sort of get down then to the tools themselves that are available, right? If I am you know building agents and I'm I'm I'm sold, you know, like this is a problem, this is a concern, I have this. What are the tools available to me?

SPEAKER_02Yeah, you know, today there you there's this big battle in the ecosystem that I'm sure Guy has talked to about somewhere. There's this battle between CLI versus MCP, right? And so we kind of have you know, you have CLI tools which very much have not don't have any really off system built into them, and then you have MCP, which the nice thing about MCP is it comes out of the box with like OAuth support. So in order to be an MCP, you have to have OAuth support, and that OAuth support um gives a pathway for saying is does this thing actually support some of the concepts that we talked about, right?

SPEAKER_02And there's there's two broad ecosystems where there's solutions that only work for MCP, and then there's some solutions that only work for CLIs. Yeah. And the question really is is that sufficient? Because what we know is there's a set of agents, computer use agents, specifically coding agents, that CLI tools are their bread and butter, right? They're just really good at it. And that makes them fast, and it's one of the reasons that like Opus and GPT-5.2, Opus 4.5, GPT 5.2 really were the were breakthrough models on these computer use things and led us to a new level of autonomy.

SPEAKER_02Um, and then you have sort of MCP only. And the question um is to think about what is the agent, what type of tools is the agent need to have access to and what type of agent should I have running in my ecosystem to do that, and how do I integrate that if it's an agent that is operating inside my company, I probably want that to integrate with my IDP. Um and I want it, but I want to think about the IDP slightly differently for agents and for users because they're actually slightly different problem statements uh broadly speaking.

SPEAKER_02And then the tools on the market tend to be um like an agentic IDP solution. You there's some tools on the market that look like an MCP gateway, which is a common topic, very much designed for MCPs, it's a giant proxy. Um, and there's other things that are designed uh specifically for solving this sort of agent-to-agent communication in a federated manner. And those are the two like landscape solutions at the moment. And then, of course, you have your traditional um last generation sort of uh vendors, like your privilege access management, that was very, very hot topic. So how do I give a human access to like a Postgres running in my Amazon or my GCP?

SPEAKER_02Very big important problem. Um, things like a cyber arc, you know, that was bought by Palo is a good example of that. Uh and those are sort of the buckets. And then of course you have your I'm building an agent that someone uses, and how does my agent or my MCP or whatever allow users to authenticate so it can carry down identity and access? And those are the broad buckets of different problems.

SPEAKER_03So like everything that you describe right now is very like enterprise solutions uh for it, right? So you in the organization you figure out how do I represent agent identities in my IDP? Yes. How do I sort of systematically control access to my systems by routing kind of a you know, blessing some sort of access grantor uh somewhere uh in it and then having it be accessed via CLI or MCP and run along? Um so like those are very like enterprise-wide. Are the solutions all like some of the problems that you've described are things that I might as an individual or as a team also say, well, within within my sort of processes that I'm running here, I want some sort of mission identity, mission permission definitions on it.

SPEAKER_03Are there solutions that are also like uh working within uh within that resolution?

SPEAKER_02I mean, certainly that's something we're trying to do with Key Card. It's trying to empower teams to be able to just quickly and easily pick up agents, but also quickly and easily build their own agents where they're not constrained, whether it's MCP or CLI or a specific scale. We're trying to build it across. And so we sort of look at it's it's actually not MCP and CLI, it's both, but it's also not PAM, it's all three combined. And how you bring these different identities, these different postures into one system that's really easy to adopt and build with is how we think about the problem, and that's where we're coming from.

SPEAKER_03And you know, I think the challenge tell us a little bit more about that, which is the how would you so if you were to use Keycard, you know, what what what does that entail? You know, if I'm uh and if you well, I guess you sort of choose, right? Whether that's sort of like if I'm uh you know the big boss kind of uh CTO kind of in the system or if I'm a developer. But you know, what what does it sort of look feel like, you know, like what am I doing?

SPEAKER_02Yeah, absolutely. So with the way that we've architected our solution is if I'm you know using cloud code and we give access to some things, I download the keycard CLI, I cut run uh claude or cursor or pie or insert your your harness of choice, keycard, we integrated with the hook system, we then you basically tell keycard, hey, this agent can access these things, and it could be a set of MCPs or it could be a set of CLI tools or tool calls on my behalf, and then we make the from that point on, no longer do you actually have to yes, no, allow always consent dialogue.

SPEAKER_02Key card can make those determinations for you, and then you get to say, you know what, if it's a delete, I should probably review. And that thing, whether that thing's running on my local device or if it's running um in a sandbox someplace or it's cloud, it works the same way across all so one insertion point to allow your to allow your agent to be evolved into a software factory and have autonomy. And then if you're on the flip side, I'm trying to build a service, I'm trying to build an agent or a tool that agents can talk to with the latest, greatest protocols without having to be a denity expert, you can use your SDKs, which then allow you to build those types of things that your agents can interact with.

SPEAKER_03And and so in that context, you are providing uh a means of some sort of central definition of logic, I guess, or sort of like the decision making. Yeah. Uh and then you are uh connecting that to different accesses. Like the two the two value propositions are one, hey, here's like a smart engine that can make good decisions around who should approach what and is able to take your input. And the second is uh, hey, I can handle the proliferation of agents. You have many, many agents who want the same kind of logic applied uh at various places, exactly.

SPEAKER_03And then I'm uh and from there connect that up to like some the parent system of you know how does a how does sort of a central platform team convey down some of those decisions into sort of individual developer desktops.

SPEAKER_02Exactly. So how does, you know, we we really are focused on how do we enable that individual developer to find lots of success, then how we let them graduate from like that to a team of people going together across independent of the harness, right? Yeah. Um access what resources and what circumstance and over time, how does that give sort of that centralized security team and the platform team both the enablement function of how do I roll this out across the entire organization, but also how does that security team get the confidence to say, you know what, we can say yes to this high high degrees of autonomy because we feel like we have the security in place that we require to do it.

SPEAKER_02Yeah.

SPEAKER_00Hey everyone, hope you're enjoying the episode so far. Our team is working really hard behind the scenes to bring you the best guests so we can have the most informative conversations about agentic development. Whether that's talking about the latest tools, the most efficient workflows, or defining best practices. But for whatever reason, many of you have yet to subscribe to the channel. If you're enjoying the podcast and want us to continue to bring you the very best content, please do us a favour and hit that subscribe button. It really does make a difference and lets us continue to improve the quality of our guests and build an even better product for you.

SPEAKER_00Alright, back to the episode.

SPEAKER_03And so cool. So I love that, and I love the sort of the dev taste of sort of a bunch of the sort of the key card uh stuff on it. And it is, you know, like I'm a believer in the bottom-up, although with the, it's not necessarily bottom-up, but in that sort of developer experience, clearly also the world has a lot of sort of uh uh uh passion for for central control systems that come along. And I guess that's the tension that sort of plays out. Neither is like right or wrong. Yeah, it's just about what is the emphasis at a different time, how much do you indeed kind of define the guardrails uh in a broad uh broad sense. But I think a key a key takeaway, uh, you know, there are many, many, many kind of interesting things in kind of all this description, and thanks for outlining that.

SPEAKER_03But this notion of these mission identities uh and uh uh uh agentically decided uh uh uh permission set for that. And then from there the ability to define those, enforce them, probably observe them, improve them over time, right? Look around them, uh enforce them, all of those are uh are probably evolutions of system over time. But that core new mission identity and permissions are sort of a uh a new entity to create.

SPEAKER_02Exactly. That's the final thing that we're focused on at Keycard is how do how do you understand what agents are doing? Yeah, how do you allow agents to do things without having to artisanally define all this strict long-lived policy? Um, and that's by we do you get there by being able to describe high-level, yeah, like you know what, these are the situations where I'm really not okay with an agent doing something. For me, it's a lot about deletion. Yeah. Um and in all other circumstances, I'm okay with a system making the judgment call on my behalf, but whether it's allowed. And that's how we can have the sort of high degree of atomic, how we can move towards high degree, higher degrees of autonomy without bad downstream things happening.

SPEAKER_02And you know, one of the biggest challenges with that talking to a lot of in the enterprise, or even individual devs, is this is great. Like conceptually it's great, but like how do I roll this out? And that's where like high-level policy combined with like the LM as a judge feature helps people to see, oh, okay, so I'm not sitting there artistically defining hundreds and hundreds and hundreds a line of no policy files, which was you know what we learned. I think a lot of security tools in the last generation is that that's where deployment went and failed.

SPEAKER_03Yeah, yeah, yeah. Because it's not like when it came to humans, we were good. Exactly. It's like a general this was like a bad problem. Uh you know, and I think the reality of uh uh of of kind of many many aspects of technology is that problems that were a nuisance, but we can kind of kind of get by with uh in the pre-agent era, but they were not great. The best teams handle them, and all those are now intolerable in the agent era.

SPEAKER_02I often I often say to people the the best practices of cloud are now base requirements for agents, and we've certainly seen those sandboxes and cloud environments and everything, and that's still true about that.

SPEAKER_03They were good before, but uh the same was true for like waterfall to to cloud. Yeah, true as well. Yes. So uh I guess kind of before we close off here, I'll ask you to do one foolish thing, which is predict the future. Uh I love it.

SPEAKER_02I love to predict the future. I do this all the time. Yeah.

SPEAKER_03So uh I mean one aspect of it is with all this identity and evolutions of it, you know, what is the transition periods to it and what is immediately carrying and all that. But if we if we leapfrog all of that and we say like three years from now, feels like a lifetime away, kind of in AI timelines. What do you think will be the sort of the state of the ecosystem when it comes to how do we define these identities?

SPEAKER_02Yeah, I think I think there's a couple things. One is um first and foremost, new protocols will actually have been deployed, right? And so right now we're really sitting on last generations like OAuth like implementations, and it turns out most of them were. Not very good, and we're seeing a lot of that when you know you go use an MCP from some of your favorite providers. You're like, well, why does this not work the way that I actually thought it would work? Why does this require all this cost? I think I think the future of the internet and and and such, agents will make everything feel pluggable, and interoperability will be incredibly important. And you'll see lots of companies really want to, because they want to have an agent forward experience, they'll really make interoperability across agent clients like a hive story, which means that like most people will be focused on actually building up net new auth components to support this.

SPEAKER_02And that will then, from an engineer perspective, make actually pulling a bunch of tools together so much easier and so much better. I think that on the flip side, what we'll see is specifically in the enterprise, we're gonna see people move where where there used to be a lot of silos, and those silos are driven by how hard it was to write policy to less silos and higher degrees of productivity and higher degrees of contact sharing. Both because like there's huge promise, but also because the identity and access systems will upgrade to the point where it allows that interoperability in the enterprise will allow these things to communicate to each other while the security team will have like the things that they need to say uh to keep the company secure, which has often been part of the reason that those systems didn't connect, is because they didn't have a good way to actually manage information sharing.

SPEAKER_02Yeah.

SPEAKER_03That's like a very optimistic view of the future. It's like, you know, the protocols will be like it's not a given that standards will be defined in three years on it uh and that sort of the system will doing it. So that is definitely sort of the aspiration, which is these systems will will evolve. And I guess, I guess kind of the the twist the uh the driver is the belief that it's just like agents are making a sufficiently enforcing function that we have no option but to solve it.

SPEAKER_02Exactly. And that's certainly been the history of identity and access for the last 30 years. It's like it's not the consumer that has really forced the evolution, it's it's excluding online shopping with TLS, it really has been the enterprise that has driven the requirement because what they want is a productivity gain. So they have the security requirements, and they go out and say to all their buyers, you really need you have to implement this. Yeah. Like for me to keep spending my 25 million dollars with you. And of course the vendors say, Well, we're not turning that money away, we really need that revenue, so we're gonna go implement it.

SPEAKER_02And also because all of our enterprise schools are asking for it. And I think what's unique about this era of agents specifically is we're seeing agents at work be adopted much faster with much higher degrees of ROI than agents at home. And you know, that's an inversion of some things we've seen previously, um, but it's certainly true now. And so I think as a result of that, that will force like the uh the uh security stack, the next generation security stack to evolve much faster than I would have typically suggested.

SPEAKER_03Yeah, I think that's a good uh like I I think I relate to that. Maybe it's like a little bit optimistic on it, but I think nothing is better, like it's a strong forcing function to do that. It is definitely well fueled from a VC uh domain of it in terms of like uh funding the transition, probably because of this sort of urgency. Exactly. Yeah. Ian, thanks a lot for coming in, you know, kind of demystifying, kind of like helping us uh understand identity uh on it. Uh looking forward to seeing you know key card and uh and in general kind of this world uh mature for it. So thanks a lot for coming in.

SPEAKER_02Thanks for having me. And if you're interested in keycard, you can find us at keycard.ai, and we'd happy to show you a demo on LPC how you can adopt coding agents and build your own agents at scale.

SPEAKER_03Very cool. Uh and thanks everybody for tuning in, and I hope you join us for the next one.

SPEAKER_00The AI Native Dev is brought to you by Tesla, the package manager for skills and context. Your hosts are Guy Pajani and me, Simon Maple. Our producer is Tom Dowler. The AI Native Dev is not just a podcast, it's a community. And we host monthly meetups at the Tesla offices in central London. Visit Tesle.io forward slash community to learn more, and I hope to see you there.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Your Software Factory Needs a Bill of Materials | Daniel Meppiel

    Nov 19, 20261 hr 1 minAudio

    Developers are about to write skills instead of code, and Microsoft's Daniel Meppiel argues the software factory that runs them needs versioning, a bill of materials and signed receipts before enterprises will trust it....

    Credit the people on this episode

  2. Paul Stack: I'll Never Write Infrastructure as Code Again

    Nov 17, 20261 hr 2 minAudio

    Paul Stack helped build Terraform. Now he says he'll never write another line of infrastructure as code, because an agent harness runs his team's production infrastructure instead. Watch all the AI DevCon NYC 2026...

    Credit the people on this episode

  3. Edith Harbaugh: Code Is Cheap. The Last Mile Isn't.

    Nov 10, 202649 minAudio

    Automated software delivery is moving faster than the people who used to approve every release. LaunchDarkly co-founder and CEO Edith Harbaugh tells Guy Podjarny why rollback should be autonomous, why one-shot coding...

    Credit the people on this episode

  4. Maggie Appleton: Agentic Coding Has Outgrown the Text Box

    Nov 3, 202652 minAudio

    Agentic coding has outgrown the text box. Maggie Appleton, staff research engineer at GitHub Next, argues that walls of CLI output and AGENTS.md files serve coding agents far better than the humans steering them. She...

    Credit the people on this episode

  5. Caleb Sima: AI Attackers Move at Machine Speed

    Oct 27, 202655 minAudio

    AI attackers now move at machine speed, while most security teams still move at meeting speed. Former Robinhood CSO Caleb Sima has seen an AI attacker up close, and he argues the only way to keep up is to take humans...

    Credit the people on this episode