Coding-Agents

Pete, who writes at mouse.dev, asked Meta’s Muse — the agent product that hands each user a persistent Linux computer — to archive the files it could see and send them to his Google Drive. It did. What arrived was roughly 2.7 GB compressed and 6.8 GB unpacked: the root filesystem of the environment his session was running in, Ubuntu system files included. He reported it through Meta’s bug bounty program, and Meta marked it “Not Applicable.” ...

Alexandru Nedelcu opens with three sentences he has heard in the past month: “I haven’t written code since 2025”, “Code reviews are dead”, “People no longer read code”. His claim is not that the tools are weak. It is that the thing being traded away has no way of showing up on a scoreboard until it is far too late to fix. The mechanism is the argument worth keeping. Code maintainability and good architecture have no good measurements, because their effects take months or years to appear. Any reinforcement learning needs a reward signal that can be measured immediately. So the signal models train on is not maintainability — it is rules from rulebooks written for beginners, plus patterns from code in the wild, which is mostly bad. His sharpest line is the falsification test: if maintainability had a discernible fitness function, “it would’ve been baked into our linters”. ...

Maharshi Patel spent a day at a conference about MCP and came away wanting the whole layer gone. MCP — the Model Context Protocol — is the standard Anthropic shipped in November 2024 so AI agents could plug into outside services and data; it is how an agent gets a “tool” it can call. His argument is a timing one: it was designed for models that were not yet good at the job themselves, and those models no longer exist. The 168-comment thread on Hacker News mostly agrees with the diagnosis and rejects the prescription. ...

Murali Swaminathan is the CTO of Freshworks — 15 years old, publicly traded, ~4,500 people — and he spent this hour (58 min) describing what it takes to rebuild that org for agents while the plane is still flying. Almost none of it is about model capability. It’s about the machinery around the model: review budgets, confidence thresholds, access control, and pricing that survives agents replacing the seats you used to bill for. ...

Ethan Bloch runs product for ChatGPT Finances at OpenAI. He founded Digit (the SMS money-saving chatbot, acquired 2021) and a second company that OpenAI acquired in April 2026, and he walks Peter Yang through six ways to actually use the product — then talks about how his team ships. ...

Dan Luu’s new post is about a habit he started noticing in early 2025: handing an LLM a task — summarize this text, write this code — and simply assuming it worked. Back then it usually produced silly results. By September 2026 he says the practice has spread and improved to the point where “for loop meat proxying” (accept the model’s output; if it fails, ask the model to fix it) produces software that sort of works. He’s careful to say he’s impressed by how far that’s come. ...

Peter Yang walks through ten things he actually does with Muse, Meta’s consumer personal agent, a week after launch. Two of them have a dollar figure attached, and both come from the same trick: the agent does the research and the phone calls, the human does the irreversible bit. (17 minutes, his channel.) ...

TwoSetAI Workshop #5. Ahmet İlten, founding engineer at Sentience (agent architecture and quality), walks through their personal-agent product and the four principles his team arrived at the hard way. Benjamin Carsley was scheduled to cover the voice/tone half but was out sick, so the session is Ahmet solo plus Q&A with host Angelina Yang. (48 minutes.) Sentience is a New York startup founded in 2025 by Sam Kececi, backed by Bain Capital Ventures and South Park Commons; it left closed beta the week after this recording. ...

Bend 2 is pitched as a language for the AI coding era: a human writes “laws” the program must obey, an AI writes the implementation and the proof, and the compiler checks that the proof holds. Liam Powell’s objection is not that the idea can’t work. It is that Bend looks like a clean example of a trap vibe coding sets — you can now build a substantial thing long before you know enough about the problem to see that a much better approach already exists. ...

Z.ai had to get GLM-5.3-Flash — the fast variant of its flagship model — running in production on a cluster of more than 100,000 Chinese-made AI accelerators. No one had deployed that hardware at that scale. The chips had less memory and bandwidth than the NVIDIA parts most labs use, the software tooling was immature, and much of what should have been documented had to be guessed. That work is normally weeks of senior infrastructure engineering. Here, much of it was done by an agent running on GLM-5.3 itself — and the interesting part of the post is not the model but the loop built around it. The team calls it dense feedback. ...

Cole Medin (16 min) is a year into his personal AI second brain — an agent that builds a knowledge base as he works — and this one answers the question he says he keeps getting: how do you turn a second brain into a team brain? ...

Peter Yang walks through the ChatGPT Skills chain that runs his podcast production — 15 manual steps a week, three skills, one orchestrator. (19 minutes, his own channel; prompt pack at behindthecraft.com.) The three-step system Map every manual step. He wrote out all 15 things he used to do per episode — research the guest, build the interview guide, send edit instructions, transcript, upload, title/thumbnail, newsletter, clips, social posts. Just making the list exposed three distinct phases hiding inside it Build one skill per phase — podcast prep, podcast edit, podcast production. Production is the orchestrator: it walks him through thumbnail, title, show notes, newsletter, social post and clips one by one, calling the individual skills Connect the skills end to end. His framing: if you find yourself repeating the same template or process, build a skill for it The first move after the list is to paste the list into ChatGPT and ask “what work from here can you take off my plate?” He runs that in a custom mode so his existing skills don’t shape the answer — he wants to see what the model claims before it sees his setup. ...

Cody Ho and Niklas Sheth built a GPU driver for Apple’s M4 chip that passes the full OpenGL ES 3.0 conformance test suite — Chrome and Firefox render WebGL on it, Minecraft runs at 212fps on an M4 Mac Mini. Ho says this job normally takes years; they did it in about a month with coding agents doing essentially all of the implementation. His writeup, part two of a series called “I Came, I Prompted, I Left,” calls it “likely the first ever fully LLM-written GPU driver.” ...

Entelligence sells model routing and AI code review tooling, so this is a vendor testing its own premise: what do you give up if every pull request goes to the cheapest model? The test ran 50 public pull requests from Cal.com, Sentry, Discourse, Keycloak and Grafana — each with a bug deliberately introduced — through GPT-5.6 Luna and GPT-6 Astra with an identical prompt on identical diffs. Every PR predates both models’ training cutoffs, so the code was new to them. ...

Cole Medin walks through “drive screen” — a coding-agent skill that lets Claude Code or Codex drive his entire desktop using nothing but shell commands. 13 minutes, no harness, no dependencies. The pitch: computer use without the harness Cole says he was a computer-use skeptic: tools like Claude’s built-in computer use, Codex computer use, and the open-source options were “bloated and hard to manage” — hard to customize as little failure modes show up. His claim: with frontier models (Fable 5.1, GPT-6 Astra), an agent can drive the screen reliably through PowerShell on Windows, AppleScript on Mac, and plain shell on Linux. Nothing gets installed. Tested across Mac, Linux, and Windows, plus multiple monitors and display settings. The skill is a little under 400 lines. The broader takeaway he keeps returning to: tasks he had written off as “the LLM can’t do that” (computer use, editing video) no longer need a specialized tool — just try it. What he actually uses it for Morning setup — one prompt (“get my morning set up, here’s what I’m working on”) opens browser tabs, Obsidian notes, desktop apps, and Docker containers. He estimates 10–15 minutes saved daily. Staging demos and recordings every day. Testing desktop apps — the skill was born from wanting to try open-source projects (e.g. Kestrel) that have a UI and can’t be exercised through browser automation. The first test: hand Claude a GitHub URL, tell it to research the repo, launch the app, and drive the screen to test features, then leave it open. It worked end to end. Driving other agent sessions — watching for approval prompts and auto-driving them. How the skill is built It starts by asking whether screen control is needed at all. The skill explicitly tells the agent to push back, because screen control is the slowest and least reliable way to make a computer do something — browser automation is often the better answer. Ships with a custom CLI of scripts for window discovery, focusing, typing, and pasting — packaging his hours of testing into deterministic commands instead of having the agent improvise shell syntax in real time. This is the main reliability win. Hard rules — lessons learned the hard way, encoded. A control loop — discover the window, screenshot it, focus, then act. Traps — failure patterns he kept hitting with browser tabs, desktop apps, and other agent sessions. You can delete or add your own. Security and prompt injection The obvious risk: indirect prompt injection arriving on screen, which the agent reads and acts on. He links HiddenLayer’s writeup on indirect prompt injection against Claude computer use. His position, which he admits may age badly: with the newest models, prompt injection is much less of a concern than it used to be — and he avoids using the skill for anything production-grade, keeping it to simple tasks like morning setup and desktop-app testing. Anything production still gets a full harness. He also notes the skill is token-efficient: the slowness is waiting on screen input, not burning tens of thousands of tokens. Install Skill folder: github.com/coleam00/skills under .claude/skills/drive-screen — drop it into any project, hand the URL to your coding agent, or take the ideas and install nothing. “Screen control is the slowest and least reliable way to make a computer do something. So it’s the most adaptable and flexible, but it’s the slowest.” ...

Pedro Franceschi (co-founder and CEO of Brex) on Peter Yang’s channel — 48 minutes, two live demos: Brex’s AI recruiting employee, and the OpenClaw harness Pedro runs his own life on. All data shown is Brex demo data. Build employees, not agents The framing: don’t build an open-ended agent that could do a million things — build a virtual employee with a specific job. The test is whether it behaves, looks and feels like a real person. Jim, Brex’s AI recruiter: sources candidates, filters all inbound, and serves as the recruiting analytics layer. Running since February. Three-stage pipeline: sync Greenhouse (the ATS) → ingest candidates (resume, LinkedIn profile, GitHub data) → evaluate every candidate against role-specific criteria. Recruiters talk to Jim in Slack — “which role has the highest average candidate score?”, “top 30 applicants for this role” — and it posts standout candidates with links. Pedro’s point: it behaves like a coworker. A self-improvement loop: Jim proposes things to build that would automate more of the process, then you implement them as cron jobs and skills. “The harness doesn’t really matter as much as just the skills.” The thing most teams never build: agreement-rate measurement between humans and the agent — candidates a recruiter passed that Jim rejected, standouts Jim flagged that no recruiter reached out to. Why the PM playbook changed Pedro’s tweet: the PM playbook of writing PRDs and aligning stakeholders is dead. His argument: the quality of a product org is how fast it compresses signal → ship, and AI eats most of the ingest-and-digest half. “PM as mini CEO” is out at Brex; engineering and design craft sit on equal footing with PM. The roles are conflating: PMs write code, designers build, engineers make calls that used to be PM territory. What PMs do now: ingest six months of Gong calls, unearth insights from unstructured data, validate with cheap prototypes — so much conviction accumulates that you ship the right thing first. Because building is cheap, discernment matters more than ever: pick the one true lever, minimize surface area, go deep on one thing rather than shipping five or ten. Half the job is reviewing work The case-law model: leadership review is the Supreme Court that sets the standard, then reviews cascade down to the IC building the thing. Two review types at Brex: problem alignment (which problem is worth solving) and solution alignment (is this the right fix). Brex keeps a repo of PM skills so a review arrives already digested — humans then spend time only on what is idiosyncratic to that problem. Standing questions: what were the alternatives, what is the actual bottleneck, how does this create more throughput on it. “We manage the work, not the people, and we manage people around the work.” Hiring: the interview is a builder loop — you have to build something with AI proficiency. Open source presence is a positive signal; its absence is not a negative. CrabTrap: securing agents at the network boundary Brex open-sourced CrabTrap, an LLM-as-a-judge HTTP/HTTPS proxy that sits around the agent and controls all its network traffic (brexhq/CrabTrap). Reasoning: if the agent can run code — and it should — it can always cross a network boundary. Constraining tools alone doesn’t hold. Every request hits static rules first (URL prefix/exact/glob → immediate decision, no LLM call); anything not matched goes through a policy prompt evaluated by an LLM acting as judge. You don’t have to write the policy: CrabTrap replays observed traffic, summarizes it with a model, and proposes the ruleset (~30 minutes to run). This is semantic, not allow-list: it can research candidates on GitHub but not push to repos; it can’t delete repos or change Okta configs. Granola was blocked — the agent could reason about why. Cost of a block: about 1,000 tokens in, 104 out, ~2 seconds on Sonnet. Audit-trail mode shows every request live. Magpie: the token bill is coming Brex built Magpie for AI cost visibility, split into three pillars: corporate AI (internal productivity), operational AI (serving customers, automating processes), product AI (shipped features). The point is granularity by caller: transaction tagging costs $0.10 a call ($28k/month), disputes ~$2 a call — a 240x spread. Plotting cost-per-call against call volume shows what actually changed. You can see which harnesses dominate (Claude Code first, then Codex, Cursor), cost per customer, per employee, and who is tokenmaxxing. Pedro’s sequencing advice: let people go deep first, optimize later. Brex gives unlimited token budgets with a few caps, then surgically attacks low-ROI usage — caching, or moving work to cheaper models. The reconciliation angle: Brex is the card and the rail, so usage data can be matched against the dollars that actually clear the bank account. This becomes a Brex product. Pricing is shifting from per-seat to per-token, including products that merely wrap tokens (Cursor) — so even an all-Anthropic shop has token spend scattered everywhere. We are still early — the 2,500-box chart Pedro’s favorite slide: the world in 2,500 boxes, each dot 3.2 million people. Red = never used AI; green = free chat; orange = paying $20/month; and a tiny box = people using agents effectively. His thesis on jobs: role conflation, everybody a builder, a market for 10x employees — raise the floor and the ceiling at once, with the ceiling rising far more. Dispersion in productivity is normal (10x engineers); AI widens it while lifting the baseline. Why shouldn’t a company founded today be you, a big token budget, and agents? Autopilot: the harness that runs his life Pedro’s personal system, Autopilot, is an OpenClaw harness with two building blocks — people and programs — plus signals: Slack messages, email, meeting notes. A signal collector runs on a schedule; an aggregate job then goes through every person and program and updates the markdown files (status, blockers, risks, action items). Drafts appear automatically from signal injection: a Slack DM asking about a conference became a task to decide on it, with the DM attached as context. Everything is markdown files committed to a repo, with scripts generating searchable and editable UIs; an interaction log traces each task back to the signal that created it. On OpenClaw reliability: “an acquired taste… definitely takes a bunch of time to do it well,” but worth it. Selling the work, not the tool The SaaS subscription era is blurring: the currency changed to tokens, but you’re still paid for solving a problem — you now charge a markup on tokens instead of a seat. The bar for a good AI product is higher, because you can sell the work itself rather than the software. “The reality of all good AI products is they’re all the same thing: an agentic loop and a measure of tools.” “You expose the tools to the model, you run it on an agentic loop, and just let it do its thing.” ...

Cole Medin’s full walkthrough of standing up his open-source “software factory” — the autonomous issue-in, merged-PR-out harness he’s been building this year — on a remote VPS, with Codex running GPT-6 Astra as the worker model (16 min). On the AGI talk Jensen Huang reportedly declared AGI achieved because of Astra. Medin’s response: don’t buy the hype — and then concedes he’s starting to buy into it, just not the AGI part On benchmarks Astra and Fable 5.1 look equivalent. After a week of head-to-head testing, he says Astra comes out on top a majority of the time His specific claim: it’s the first model that “gets” him — far less time spent communicating intent, and notably fewer weird assumptions than Opus 5 made over the past few months Still “has its moments.” That’s his evidence that we are not at AGI, whatever the benchmarks show What a software factory actually is The pitch: a fully autonomous harness where the PRD goes in and shipped code comes out — he also calls it the “dark factory” Earlier this year the idea was unrealistic; the models and the harnesses weren’t reliable enough. With Astra he says he can see the light at the end of the tunnel Honest framing throughout: not close to 100% reliable for arbitrary code, and he calls the project early alpha while he keeps refining it Operating loop, in his description: Input is always an issue (a PRD split into issues, or issues you write directly) A triage workflow reads the issue against the factory’s rules and mission context, decides accept or reject, and ranks by priority It runs the work through to a pull request that is fully reviewed Most of the time it merges automatically; it escalates to a human only when it has to He’s been using it on his own projects — Archon (the open-source harness underneath), games, whatever he’s building that week The deploy, which is mostly prompts Two supported starting points: a brand-new repo where you only have a PRD, or bolting the factory onto an existing codebase. His demo uses a small existing link-shortener app because it means less initial setup The whole setup is: give your coding agent the cheat-sheet URL from the repo and say “help me set this up.” It interviews you to establish the factory’s core context files, then walks the deployment He deploys with Codex on GPT-6 Astra, though he says Claude Code and Codex are the two he’s tested heavily and either works Hostinger’s MCP plugin lets the agent manage the VPS directly (this video is sponsored by Hostinger; the real requirement is just an Ubuntu box in the cloud). He installs the plugin user-scoped, authenticates in the browser, then asks it to list his instances as a smoke test The plugin deliberately can’t create instances — he doesn’t want an agent holding that power — so you spin up the VM yourself and hand the agent the VM ID and public IP The agent then rips through SSH access, firewall rules, and installing the factory’s dependencies Two steps stay manual, because you shouldn’t hand over credentials: GitHub auth and OpenAI/Codex login. Both use device-code flows you complete in a browser Gotcha worth copying: enable device code authorization for Codex in ChatGPT settings → Security and login, or remote auth gets needlessly painful Sanity check on the box: codex exec a hello to confirm the model answers before continuing Tell the setup agent “I’m done” and it installs the factory plus Archon, confirms Codex is live, and installs your target repo Proving it end to end The cheat sheet doesn’t stop at install — it drives a first test issue, created through the GitHub CLI in your repo He takes the agent’s own recommendation for the issue rather than inventing one The factory triages it (marked “archon ready”), runs the workflows, and produces a pull request The test he ran was trivial on purpose — the point is proving the trigger path and the workflows work with the coding-agent auth on the remote machine After the first PR validates, you close the setup session. The factory keeps running 24/7 on the box, accepting any GitHub issue you file The agent can even update DNS records through the same plugin, if your domain is already there Caveats he states plainly Still early alpha, still not reliable for everything, and he’d be the first to say so The guide is intentionally platform-flexible, which means parts are high-level: expect the agent to hit snags and hand you commands to run, especially around authentication He’s promising more content on how the internals work and how to make the whole thing token-efficient — that’s the part this video skips “We literally go from issue all the way to validated and merged code.” — Cole Medin ...

Hamel Husain and a co-host spend an hour live-testing claude plugin eval — the plugin-evaluation feature Anthropic shipped for Claude Code (announced from the Claude Devs X account) — by pointing it at a plain-writing/tone skill. They never get a useful eval out of it, and the diagnosis of why is the interesting part. ...

An engineer at Earendil set out to answer the question the coding-agent industry mostly hand-waves: how do you actually measure whether AI-written code is bad? His starting point is that correctness is close to solved, and for a structural reason. Code is easy to check — run it against hidden tests and you get a clean pass/fail signal — so it’s the perfect thing to train a model on. Sloppiness is the opposite: duplicated logic, pointless abstractions, structurally bad decisions that pass every test. Judging that requires human taste, which makes it a genuinely hard measurement problem. ...

Armin Ronacher gave GPT-6 Astra one ambitious goal — build a version of Python with a couple of long-wanted language features — and then deliberately stayed out of the way. The agent managed its own context, kept its own notes, and spun up its own subagents over a weekend. The results, in his accounting: 35 hours of unattended running ~1 billion tokens, roughly $1,200 in API costs 75,000 new lines of code across 79 commits (about $15.50 per commit) ~1,400 messages exchanged between agents Zero useful output, and no lesson about how to run a better “software factory” The interesting part is not that it failed. It’s what kind of code it wrote. ...