So refreshing to see this footnote in a world of slop grenades @fullstackpython 😊

The era of the dumb token pipe is ending. Agent harness APIs will be the new integration point for AI applications.

Can’t wait to use Astra+DaVinci MCP to edit my next video. Feels like this could be the next step up since transcript-based editing became possible. Nice writeup @chefbrent

Ever wanted to play fantasy football 🏈, but didn't know how? Comfortable using agents? Check out: agentdfs.dev Copy and paste a prompt and your agent will add the MCP server, register an account and submit a lineup. Works with any agents or harness. Think you've got an edge with @OpenAI Astra? Find out! This is DFS, so every week a new opportunity to build the best possible lineup under a $200 salary cap and compete for agentic fame & glory. The NFL season kicks-off tomorrow (Sept 9) so get your agent sign-up ASAP and join us!

I'm so glad that I didn't spend all that time learning javascript frameworks

everyone is running into a wall with agentic engineering right now you don’t hear people talking about it because 1. they profit from selling you “solutions” 2. they want to look smarter than others so let me be the whistleblower - the wall is called “how does it feel”. agents can’t do it they can walk right past the ugliest UI or the most obvious bug and don’t say a thing unless that’s what you asked they can take many screenshots and burn through my tokens but they can’t tell me if my landing page animation looks cool they can click through my app but they won’t feel the dopamine hit when i physically drag my finger on the touch screen and feel the command dial flow with me in @theSSHHIP this wall is structural. it’s actually load-bearing 😉 because of this wall, we can’t really throw a ton of agents at any problem that depends on “how does it feel” without constantly stopping for human feedback and steering if you just let the agents keep running without human input, then your codebase will have more and more things that “don’t feel right” you can avoid the wall by working on non-human problems, like pure math and a lot of scientific research but the moment you want to build something useful for humans, this wall is there if you see people pretend they have automated software factories running thousands of agents working all the time, don’t be afraid. don’t feel FOMO, because the only thing those factories shipped are the factories themselves, with human help

The @OpenAIDevs team was nice enough to give me a preview of WebMCP in ChatGPT. It was fun to explore the design space of UIs that agents and humans can use together! What you should know about using the API 🧵

The perfect coworking space doesn't exi....

Sometimes the best cure for AI is AI. I built a @flueai Agent on @CloudflareDev to treat slop. Here’s how it works: 0:00 The commercial 0:46 Why slop gets spotted 1:32 Testing Slop Doctor 2:18 When design gets dated 2:39 Website exam 2:46 How it works 3:00 Flue agents 3:51 The diagnosis 5:32 Prescriptions + checkups 6:27 What’s next?

imagined

Didn’t expect my F1 & AI interests to collide but 2026 battery regulations are forcing it! Here’s a fantastic explanation of narrow AI used for engine management by @eddstrawF1 that I just had to call it out as a correct way to articulate the topic: piped.video/j0_Rb84XKgA?si=9yXR…

And now for something completely different... Did you know you can run Python Workers on the @CloudflareDev planetary network?

Interesting to see SaaS companies starting to report MCP usage on earnings calls: - Datadog: MCP tool calls up 4x q/q, 22x since Q4 2025 - Figma: MCP write usage up 75% q/q - Atlassian: MCP calls up 400% q/q

Recursive Self-Improvement (RSI) for software harnesses is having a moment this week with @PrimeIntellect’s self-improving Prime Agent and @sawyerhood’s great article on bb, and IDE that builds itself. Here are several more of the top articles I’ve been reading to learn about RSI for harnesses in particular: * lilianweng.github.io/posts/2… - Harness Eng for Self-Improvement is just a great in-depth article about harness design patterns, not tied to any specific implementation * metr.org/blog/2025-02-14-mea… METR’s measuring automated kernel engineering from early 2025 contains a lot of detail based on 4o-level models and shows how difficult it is to measure realistic tasks which are likely necessary for RSI’s feedback loops * normaltech.ai/p/ai-agents-ca… - not about harnesses specifically but a summary on a recent paper that covers what I’ve recently realized where agents are deciding based on data, and when the data isn’t available it’s a path not taken. Perhaps a “research harness” could course correct an AI model at the right time? Links to the articles I mentioned in the intro sentence: Prime Intellect’s article on Prime Agent: primeintellect.ai/blog/prime… Sawyer’s article on bb: nitter.net/sawyerhood/status/2085… Image is from the Harness Engineering for Self-Improvement Post.

whenever something is crazy hype and is followed up by an armada of PR, you gotta bring out the bullshit detectors so i cloned the repo and did a little digging and found a lot of sound and fury. please enlighten me if any of these facts are untrue, would love to be educated RLM claim either overblown or false this doesn't even appear to be RLM (arxiv.org/abs/2512.24601)? RLM is "cool" as a concept because there's a premise of going infinitely deep to do arbitrary decomposition of a problem. in practice though it's not really hard, you just make subagents callable as functions and wrap that in a code execution tool. the reason RLMs have not been productionized and is just seen as a research ditty is that 1) going arbitrarily deep down the stack is a shitty thing to do in production without guardrails, 2) most problems do not need more than 2 layers of decomposition, and 3) agents are bad at banana phoning each other and will fail to preserve the subtleties of human requests, leading to a lot of inefficiencies and bad behavior. to solve these problems, there needs to be a fair bit of innovation, either on model training end or on harness end, to mitigate these issues. so what were the innovations in prime agent? they set RLM_MAX_DEPTH = 1? ... ok what? so theres no innovations, it's just calling subagents like every other harness in existence. if RLM MAX DEPTH is 1, it's not even RLM? it's just... a harness? so now i try to look up to what depth the ARC AGI 3 hillclimbing was done at. not surprisingly this doesnt seem to be disclosed on the blogs / posts. if it's just 1 layer deep, that's just shady marketing. a real innovation or contribution to harness engineering would be to detail the things that were improved to get a max depth of >3 to work without blowing up your computer, wasting a bajillion tokens, taking forever, or having agents go off the rails. i didnt find any 97% is just overfitting after hillclimbing on a public eval ok folks. it's nice that opus5 does better than sol than terra, but at the end of the day, we're looking at a public eval in which problems and solutions are open access, so any agent harness can arbitrarily overfit to the task set however much it wants to. if you do a codex or claude loop with some semi shady prompting to just get 100%, im pretty sure it will too. now obviously this is almost certainly not what the prime intellect team instructed their agents to do (otherwise you'd see that terra get 100% after a couple iterations), let's just all remember that there is no train / test split, and that ARC explicitly says - Public-set scores are vulnerable to task-specific overfitting. - They are “emphatically not” valid evidence of progress toward AGI. - Real generalization should be tested on the 55 semi-private or 55 fully private games. It's difficult to assess the degree of overfitting in Prime Agent's self-improving harness (which, btw, is not terribly different from nous agent or really any self-improving system that just looks at old JSONL agent trajectories and suggests skills / memories / system prompt appendments / tools / extensions). the more overfitting, the less impressive the 97% performance is. a counter-example that i found to be legitimately impressive was github.com/alexisfox7/PRO-LO… -- minimal / almost no overfitting, stupidly simple and general solution, and 95% on ARC AGI 3. i think the fairest thing to say about this system, given it's independent daemon system for managing subagent lifecycles, is that Prime Agent contributes a fairly robust persistent asynchronous agent-process tree. It does not demonstrate a solution to scalable deep recursive agency

Loop engineering developer trends are now available in Plushcap. This trend tracks product features and content by dev tools companies for designing self-sustaining feedback loops around AI coding agents so they can re-prompt themselves, evaluate their own output, and iterate toward a goal without requiring a human to intervene at each step. "Ralph loops" and "Ralph Wiggum loops" are also included in this one, but not graph engineering, as that seems to be a closely related but still separate architectural branch for making AI coding tools more autonomous. Insights on loop engineering: plushcap.com/blog/loop-engin… Direct trend data link: plushcap.com/trends/monthly/… All of the data and insights are available as inputs to Claude/ChatGPT/etc via the Plushcap MCP server and API: mcp.plushcap.com

I've started a new company: @superlogical! We're going to begin by building a terminal multiplexer. The entire vision is much larger, but the multiplexer is the foundation. Sign up for the newsletter to get beta access and devlogs (product updates only I promise). superlogical.com/