Early Days Of Agent Recursive Self Improvement | Node 7.15 Upgrade
28 minAudio
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via their podcast app and to have them all in one place. If you'd like to participate they are held weekly on the @MorpheusAIs Twitter spaces. So enjoy Episode #95! Recorded: 15 October 2026 AI SUMMARY: In this week's town hall the market backdrop was noted while highlighting what matters more: the network keeps building. Despite broader volatility Morpheus throughput has pushed past 60 billion tokens processed with activity still trending strongly. Model usage across the network shows DeepSeek v4 Flash still leading and GLM 5.3 Flash rapidly gaining ground. Alan walked through the 7.15 node release which includes important security fixes stronger AMD SEV quote validation tighter delegated API permissions and new provider capability profiles to improve how reasoning settings are handled across different back ends. The message was clear: if you're running a provider node upgrade to 7.15. The series on building effective agents continued focusing on recursive self-improvement. David explained how once you've set up your agent second brain and coordinator loop you can dedicate part of your worker capacity to improving the system itself by capturing daily learnings turning them into proposals and having agents refine their own workflows over time. Practical challenges include feeding in new tasks avoiding context bloat and managing session timing when connecting directly to the Morpheus network. The familiar theme for this community remains: price moves come and go but real utility active contributors and working agents are what keep the momentum going. https://www.youtube.com/watch?v=0u3anxjNA44 Timeline: (00:00:00) Intro (00:00:30) Market dip, morale check and network throughput (00:03:37) Node 7.15 release with security and provider upgrades (00:08:22) Building agents with recursive self improvement (00:16:52) The next bottlenecks: task intake and context limits (00:19:30) Managing Morpheus sessions, staking and Cron workflows (00:23:29) New open models, Tesla speculation and community outlook If you would like to support me for the value I provide you can use the referral link below when depositing capital and we will both earn bonus MOR. Alternatively you could send me a token of appreciation directly to my wallet ;) Referral Link: https://dashboard.mor.org/capital?referrer=0xE935f231c99c04Ee0b4532a3d0BdA81B152a0384 Wallet Address: 0xE935f231c99c04Ee0b4532a3d0BdA81B152a0384 Connect with Mere Mortals: Website: https://www.meremortalspodcasts.com/ Discord: https://discord.gg/dX9BEhT3kz Twitter/X: https://twitter.com/meremortalspods Instagram: https://www.instagram.com/meremortalspodcasts/ TikTok: https://www.tiktok.com/@meremortalspodcasts Value 4 Value Support: Website: https://www.meremortalspodcasts.com/support Paypal: https://www.paypal.com/paypalme/meremortalspodcast
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.
Bowen BelmerWe are we are live. It is the weekly town hall for the Morpheus community. It is Thursday, 10/15/2026. Let me just make sure we got pushed live. We did. Okay. We got pushed live on X and YouTube. So I'm obviously not in my normal setup. I'm traveling at the moment. You know what's funny is my audio and video on my laptop is, like, arguably just as good as my, like, actual desktop setup, which is really depressing when I when I see it play out. But anyways,
you guys, how how is everyone doing? Everyone hanging in there? Doing Doing great. Thanks. Doing well. Good? Thanks, Bo. I I have to say,
David Johnstonwhen you just said people are already calling it, it reminded me so much of that scene from Civil War where the president is practicing the speech, he's like, people are already calling it the greatest battle of this.
Bowen BelmerYeah. People are already calling our comeback the greatest comeback of all time, so that's what I heard.
David JohnstonYeah. I love it. People are saying. People are saying. Exactly. It's coming here. Yeah. It's one of those things why, like once said.
Bowen BelmerYeah. And that man was me. No. Yeah. No. It's, you know, look. It's one of those days. Whole market took a took a huge hit yesterday. Obviously, equities across the world did really bad. So it's... You know, it was it was kinda just bad across the board, but we persist. The building doesn't stop. And just taking a look right off the bat, mean, we kinda start with this every week, but it is, particularly worth looking at this week because of, you know, the fact that things are things are struggling in in terms of market and token price is the throughput stats.
So despite the fact... And let me, we're gonna look at native. Despite that fact, if we look at throughput, we're over 60,000,000,000 tokens processed now and continue to be mostly up into the right. I mean, we're at the exact halfway point, for October, which would basically imply that we're maybe slightly behind schedule where we were in September, but, you know, there's still time for, obviously, the month to finish strong, continue that up into the right trend.
I would say even if we're on par or better than August, you know, I don't think it has to be the case. If our benchmark is every single month ever needs to be the best month ever, obviously, that's great, but also, you know, not necessarily gonna happen. But either way, you know, it's great to see continued throughput, continued activity. I know I've still been fairly active. I know you guys have as well. Do you guys have any anything on on this specifically or on, you know, throughput or the models? DeepSeax, v four flash still continues to be the dominant dominant model here.
David JohnstonGLM 5.3 finally sort of had a breakthrough. It feels like people were sitting on five two for a while. But if I look at the numbers, it seems like five three flash is now... I mean, it's neck and neck now for DeepSeek v four. So clearly, you know, it's it's on the verge of of surpassing it. For me, it's it's the best model out there today. And I have to say, you know, I appreciate, Alan, all of the security upgrades you guys have met... Made. I feel a lot better about putting
more and more to work when it comes to staking and using the p two p node with all the security upgrades that have been been made. So Yep. Yeah. Tip of the hat to to you and Alex and everybody else has been working on that.
Alan (Nomadic Rogues)Yeah. We'll talk about that in a second. But, yeah, that was was a good find.
Bowen BelmerYeah. I guess to that end, was gonna say, Alan, Kyle, any any news from a protocol perspective or just any updates, you know, from from code within the community this week?
Alan (Nomadic Rogues)Yeah. I'll I'll talk about two. Yesterday, we released seven dot 15 of the node. And there are two big things that I think matter to the to the population. First, security. We've... Community member found some issues that were responsibly disclosed. A real good big thank you to the community for that. It's always difficult. We've got a lot of folks that are finding things and routing them to us and so chewing through them and, you know, what's real, what's not real.
And these were a couple things that that were certainly affecting the environment. So things like a delegated API user can no longer replace the operator login or grant itself permissions. It does not already have. There was a there was a permission bypass. IPFS downloads and uploads have been limited to stay in the full... Inside the folders that you set in the configuration. And on our AMD SEV quote that does not match the measurements is now rejected. We were previously an empty measurement was skipped and so a bad quote could pass. And that affects our ability to say, look, TEE
environments are completely trusted. The second big win, and this was also a community... A customer ask, was what we'll call provider capability profiles. You know, so every chat model on the health check, if you're on seven fifteen, now says kind of which back end is serving it, whether reasoning can be turned on or off and which parameter the back end actually accepts. And you can declare that in the model's config as a provider or leave it blank and the node detects it.
So the hosted API uses those profiles to help steer and give some some deeper functionality to your request, whether you want reasoning on or off. And the difficulty here is, if you think about it in a distributed peer to peer, the way people present providers of LLM models are gonna be different. Some could go through open routers, some could go through light LLM, some could be native through the proxy router.
And how those things are configured for things as simple as reasoning, thinking on or off, they call it different things. Right? So open router... I I can't remember exactly what it is, but they call it one thing. Light LLM calls it something else. The OpenAO protocol standard calls it something else. And so when you send something through the API gateway and you say, look, I don't want you to think about this, turn reasoning or thinking off, controlling that over the permutations of provider environment and model is
pretty challenging. So this was the first set of reasoning and we will have some others that come through that allows us to for better control. Now the critical piece, and this is for all you providers out there, please upgrade to the latest seven fifteen because that will bring this forward. And so that's really important to make sure that we're on par. And again, another challenge with a distributed network is we don't have direct control. I can't go out and beat people over the head and say, hey, upgrade your provider node. So I'm begging,
I'll beg that all the providers please upgrade to the latest seven dot 15 because then it helps the network. You've got the latest security patches, You've got the latest functionality to participate on the network, and it's all good. So seven dot 15 is out. Please go upgrade. And the release notes, as always, are in the GitHub repo.
David JohnstonWell, also, I can can say for myself, I upgraded this morning. So... Excellent.
Bowen BelmerGo ahead, Will. I'll call some people, Alan. When Alan says do something, you guys should do it. Guy knows what he's talking about.
Alan (Nomadic Rogues)I heard it from a guy who wrote from a friend.
Bowen BelmerYeah. Yeah. Don't listen to me saying you should do it. Alan told me you should do it. No.
Kyle (Bow Tied Blue Fin)That's awesome.... Standpoint. Right? You're you wanna update so that you have access to people who are filtering by version number. From a consumer update, you want prod... Providers to update so that you have a bigger set of people you you could, you know, route your request to. So whenever a new release comes out, it's in everyone's best interest to update to it, whether it's your security functionality or or or scale.
Bowen BelmerYep. Awesome. Well, thank you guys. That is great news. Everyone, go update. Yeah. That was all, I guess, from the the protocol side this week. David, I know we got a couple avenues here. We can either jump into new stuff or go, into the recursive self improvement stuff. I think we should probably continue that thread first.
David JohnstonLet's, let's do it. So for those that have been tuning in the last few weeks, we covered how to set up your agent, then we covered how to set up the brain, your second brain, so that it really understands your context, how you can build your standard operating procedures or skills, and really sort of index all of your long term memory. And so once we got through, you know, sort of both of those steps, you know, we talked last week about looping. Right? So, okay, you've got an agent,
but now how do you make sure it's it's getting the work done? Having a coordinator agent is a real big breakthrough. And it could be as simple as setting up a cron job that wakes up every 30 minutes and checks on, what the worker agents are doing, and sort of, pushes them forward. So those are three powerful pieces. The agent, the brain, using the brain to enhance the prompts of what your workers are doing and then a coordinator to to follow all of them. Once you've got that set up,
you're now in a position to set up a recurring self improvement agent. And I did this in the last month, it was a real breakthrough for me. And it can be as simple as telling your agent, hey, I want to take all the learnings each day and put them into proposals that get coded up by the workers to improve the process itself, right? And for me,
recursive self improvement agent has been able to come up with innumerable ways to improve the way that loop works, to improve the way I'm you know, taking the learnings themselves and applying them. And so you know, I'm effectively dedicating almost 50% of my worker time now to improving my workers. Right? And improving that loop, improving the architecture, improving, you know, you can really take new things you learn every day,
to improve the cadence of the workers. Sometimes there's a lot of idle time going on in the loop. And just things that are sort of hard to see from the outside observing, having that learning be applied every day has hugely accelerated the outcomes. So my agents went from a few things done every day to tens of things to, you know, I think I'm north of a 170, tasks completed. And and so, you know, once you've got this all working together, really the last bottleneck becomes, you know, the ability to,
articulate to your agent what you want to accomplish. And that sort of the ultimate human gate is you acting as an oracle into the world and saying what you wanna do and adding those tasks. Right? So, you know, I I'd really encourage people if you've set up this type of system to add the RSI agent to the process and, know, you can get really efficient, right, compared to where you started. And now you sort of have this this self improving loop going. So, Bo, I don't know if you wanna ask any questions about that. Yeah. Is is... No.
Bowen BelmerI think that's a good overview. Thanks, David. Is is that RSI agent, you said that that's a skill on Bernardo's GitHub. So if I point my main agent towards Bernardo, it'll be able pick it up. I've put a lot of the skills there.
David JohnstonAnd basically, what I do is I take my standard operating procedure skill, and I have Bernardo build out a version where it's sort of generalized, right, of taking all the personal data out, and I've released those one by one. I'll go ahead and and update the RSI skill that I've I've built internally and release that today. So folks that listen to this, I will be able to download it. But, yeah, they'll all be on that same Bernardo public GitHub repo.
Bowen BelmerOkay. Awesome. Yeah. Because, I mean, that's generally speaking just like how I experience agentic AI right now, and I think a lot of people do is just pointing my agent at a repo and then asking questions about it and, you know, having it do a lot of the thinking for the things that I I do not understand. So I suspect others probably don't either. That is really helpful, though. And, yeah, I mean, this is ultimately the goal. Right? It's to have have this assistant who understands your intent
or assisting or partner, however you wanna view it, does the things that you want it to do, learns the lessons so you're not... Because this is something that I find myself doing quite a bit right now is, like, asking my agent to do something. It'll do it or it'll kind to do it, then I'll have to be like, no. We already talked about this. Like, don't do it this way. Do it that way. Oh, yeah. Sorry. You're right. Looking back, we did talk about that as opposed to just teaching itself the lesson. It's like it's like fifty first dates sometimes.
Alan (Nomadic Rogues)You know? So I Oh, it's Yeah. It's like it's like the kids. I've I've beaten this drum before. It's it's a... You know, a couple episodes ago, I talked about kids. Right? That's reinforcement learning. And and just like that, no, we talked about this. You're not allowed to go and and and play in the street. That's... That would be bad. You know, no, you're not allowed to use my wallet to go to go spend money. Now with the RSI, I think of it kind of like these these are teams. If anyone's been, you know, near a corporation or enterprise, they've got these... Some of the Japanese concept of a kaizen.
You just take one process and improve it a little bit. And that's what you're doing. I'm like, hey, this was great. What did you remember from today? What can you improve tomorrow? And you end up spending, you know, I think to David's point, you end up spending x percentage on that, not the work that you wanna get done, but improving the future work, and it pays dividends In teams of people, in corporations and other groups, doing a kaizen, which is basically look at what you're doing, do what you say, say what you do, and then improve it
helps teams behave better. And so if you take that human analogy or analog for growing kids, raising kids, raising teams, developing teams, We're doing the same things with our bots and our our agents is helping them get better. And sometimes you gotta spend that time and those tokens on, I'm not gonna do a task today. We're gonna learn how to do next... Tomorrow's task better.
David JohnstonMhmm.
Bowen BelmerYeah.
David JohnstonYeah. I love that I love that comparison. Yeah. So I mean, we're we're sort of rebuilding the best practices we know that work and it's just accelerating through this in in a fraction of the time. Right? The
Alan (Nomadic Rogues)best the best the best part is the the agents don't argue. I don't wanna do that. I don't have enough time. I... Oh, why is this self improvement so hard? Agents are like, yeah. Got it. Check. Go. Wouldn't... Yep. Don't argue.
David JohnstonMaybe maybe next time, and I think this is something I wanna explore too, Bowen, is we can talk a little bit about how to feed new task in because this is not that I'm I'm working to crack is to better, I I think there's a couple improvements I still need to make. Now they've got an agent, it's got a brain, it's got looping, it's got RSI, there's still... There's always a bottleneck. Right? And I think the next bottleneck that I'm that I'm trying to resolve is one,
how to how to better feed in items that need to be worked on. Right? So I've connected it to GitHub and it's pulling down items. But then, really, the tough thing is what's what's ready for production. Right? What looks like... And how to have that holistic approach. Right? And I... And I've taken some some learning agents and and tried to address that, but I think that might be a good topic for for next time because once you get this engine going, it's like, Now how do I pull more stuff in for it to do? Becomes the bottleneck.
Bowen BelmerYeah. No. That would be helpful.
Kyle (Bow Tied Blue Fin)I think that pairs equally with another bottleneck on context size. Right? Because as you start bringing in all your prior learnings and as you start bringing in, you know, all the different things that you're working on and hey, what do I need to work on next? You start to build this hundreds of tens or hundreds of thousands of tokens of context then Mhmm. You know, that it it costs more
than what it looks like at face value. Most people, when I just said cost, probably thought of dollars. Right? But it's not just dollars, It's intelligence. Right? When you have massive context windows, you end up with the needle in a haystack problem where you've just, you know, injected all of these things for the LLM to think about where the first sentence rock solid. It knows exactly what you're gonna say. It's gonna follow the instructions perfectly. The last sentence,
same thing. It's gonna follow those instructions perfectly. But once you start to talk about this, you know, the second fifth and the the third, you know, the fourth fifth, it's like the... They... Just like how you read sometimes these LLM scan a little bit and and lose some of the granularity and some of the details in these ultra long context windows. So I think that as you do recursive self improvement, as you bring,
excuse me, more instructions and more valuable things that you wanna know about into the prompt. You just as much have to think about how do I scrape out all the stuff that I don't wanna talk about right now because you're you're clouding, you know, that context.
David JohnstonYep. Plenty more challenges to cover. So, yeah, let's let's keep the series going, Bo. I think it's really... Yeah.
Bowen BelmerYep. We definitely will and also, just so guys know too, like, we're working on other, like, visuals and things to show in practice. It's... Some of this stuff's hard, and we talked about this the last two weeks, but it's kinda like the equivalent of telling somebody, hey. Look at this code. This is how this works. Okay. You saw it? Cool. You know? As... Unless if you have, like, live tasks and things, to show. So that's very much something I'm paying attention to too and trying to figure out, because from a viewer perspective, obviously, having something to, like, look at and experience is great.
I do have another, I guess, topic that I think is worth talking about, and maybe this is another one for another week too. But it's just something that I've experienced, with my own agent as I've kind of transitioned from... At one point, I was exclusively using the API gateway. Now I am partially using the API gateway, partially having my agent, you know, connect directly to the, the Morpheus network. When I do that, I'm finding that I'm having to pay attention to, like, session management.
You know, like, if I if I text my agent and all of a sudden it opens a session and it burns through... And part of this is the way the more tokenomics work currently, right, which is where you're, you know, opening a session for a period of time. And there's there's scenarios where I'm like, oh, I actually wanted to do this other thing today. Oh, shoot. I still have... I don't have enough more. Like, I have it staked, and it's gonna get freed up this evening, and then I can do that thing. And so learning how to, like, manage your your session opening and and,
I guess, the timing around that based on what it is you wanna do, I think, is kind of an interesting topic. Because for a lot of people, they're gonna get kinda bare minimum more to get going, you know, like newer people especially.
Alan (Nomadic Rogues)I've I've started I've I've, started with my, my GrokBot harness that I I I demoed is when they use the Morpheus inference is I have them wait a little bit, so open a small session and then close it because you get the unused more back and that's helped that utilization is you actually tell the harness, hey, open the session. It takes a little bit of time, but if you're gonna use it and you you have a couple turns, then it's still a warm session. But once we're kinda done, close it down rather than having it stay open because you're right, you'll you'll lose... You'll you'll
lock up more more and you're not using it. So I see a lot through the uplink gateway as opens and closes and open closes. So the payment in that is base ETH. It's the small open and closes on the on the ETH, which isn't horrible. But I'd rather pay that than lock up more for a session for a session time that I'm actually not prompting. So that's a... You know, you... Your harness can... With the Morpheus skill, your harness can go, oh, yeah. Open, do a couple turns, and then close it down. And that's helped me.
Kyle (Bow Tied Blue Fin)Yep. Very cool. I think what Alan said is really important. Right? That there is a natural trade off on how short do you wanna create a session, but how many times do you wanna pay the toll to to open that session. And I think it's, you know, Uplink is an amazing tool. It helps manage that session open and close and it it makes it simpler both so that that problem you said you don't have to remember to go close the session. You can kind of set it to close that session for you, but I just recommend that that people really think more about this. If you're opening a session to stay open for three seconds and sending like a test message to a
GLM flash that's gonna cost you 0.001¢ and your transaction cost was 0.04¢. You know, it it it like keep... You can keep these open for a little for a little while depending on what your activity is. So, like, there's a there's a good optimization and a good tuning there.
Bowen BelmerI mean, I think Alan nailed it on on that comment. So keep keep that in... Yeah. That's sort of, I guess, how I've been doing it so far, but then also for cron jobs, trying to figure out, like, exactly how long it'll take and leaving just a little bit of a buffer. Like, the, like, custom newspaper I have my agent make for me every morning, I know, like, basically how long it takes so I can open a session, know, slightly longer than that to make sure that it, you know, has the inference necessary. Yeah. That makes sense, though, what what you guys were saying. And that's, I guess, kind of in practice
what I ended up doing based on my agent's suggestion of, like, reading the docs and saying, oh, this is probably how you should do it. Very
David Johnstoncool. Sounds good. Well, I know we're, we're wrapping up a bit early with everybody, traveling, but just wanna call out that another set of amazing open source models got dropped. Muse is the hot thing, this week, and just, like, the drumbeat continues to accelerate. It's hard to believe that it was only, like, two, three weeks ago when everybody was talking about pausing, and then giant series of releases immediately came out. And it's like, I I think everybody's just completely forgotten that was, like, you know, two weeks ago or whatever, three weeks. Only twenty five minutes ago, Zuck released,
Bowen BelmerMuse Spark one open weights releasing today.
David JohnstonWhat do you know? Perfect. There you go. I mean, every time. Good times. Well Yeah. Perfect timing. So A key.
Bowen BelmerYeah. Another entrant, into the space, which is great. It's... Yeah, it is funny to me how we just, oscillate back and forth between we're all gonna die and pedal to the floor accelerate. You know? It's like there seems to be no no in between, at least not in terms of, like, public zeitgeist.
David JohnstonYeah. Well, whenever they need to fearmonger, they they pull out the one. And, you know, in the meantime, they like making lots of money, and the free market is doing its thing. So we'll get, acceleration in reality. So Yeah. Hopefully, like Oh, wow. So
Bowen BelmerWell, the only other thing we're saying is if you're listening to this, you're obviously a tech nerd. So don't forget to watch the Tesla Roadster demo this evening, which is gonna be... I I don't know what it's gonna be, but I'm excited about it.
David JohnstonSo we'll see. I I hope it flies. That's that's that's mean figured out some deal with the FAA and there's no longer a 400 pound limit on, you know, the need for a pilot and all of a sudden it's like we have flying cars. That that would be amazing. Yeah. Full self drive in three dimensions.
Bowen BelmerExactly. I was looking at this. There's actually a poly market for this.
David JohnstonLet's see. Of course there is. Of course there Well,
Bowen Belmerthe big news there is are SpaceX and Tesla gonna merge because there's rumors of that and there's rumors that this announcement is gonna be, you know, something to do with it flying and then partially announcing the merger. But right now, the poly market for will the Tesla Roadster fly for more than one second, It doesn't have a lot of volume. We're only $45,000 in volume, but 74% of the money says yes.
Kyle (Bow Tied Blue Fin)Wow. I feel like Bo might have an advertisement deal here with the Yeah. It's like in the back. Pew back every time.
Bowen BelmerNo. What's funny is, you know, I'm I'm historically not even a fan of the actual Tesla cars even though I acknowledge that they're amazing tech. But this is just, like, objectively so cool. I can't deny it. You know? It is it is incredibly cool. And, yeah, no no prediction market sponsorship yet, but if you're listening, Polymarket and Calci, we could we could talk. Good stuff. Yeah. But, no, everyone hang in there. You know, it's funny with... I've... Morpheus the whole time has been just the ultimate, like,
community signal, high noise, low projects, and it's persisted through all price volatility, and it just continues to persist. Like, there's a really core group of people who care about the vision, care about the tech, you know, continue to contribute, and, it's awesome. And we, you know, we'll live to fight another day and keep keep pushing onward. Keep moving. The good thing about a crypto project that I shouldn't even say crypto project. A project that's adjacent to crypto that has an actual use case is regardless of the fact that price action was trashed yesterday, guess what? My agent still works today. My newspaper was still delivered this morning. You know? So it's like, we... Onward.
David JohnstonAwesome. Thanks, everybody. Appreciate it. Thanks, Bo.
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...
Zion, hear me! For Season 2 of the Mere Morpheus podcast I'm going to be uploading the Morpheus builder weekly updates. My reason for doing this is to make it more accessible for those who want an update directly via...