We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Stop building features you can ask for
In a new video, Theo argues that engineers under-use agents because they "think in code". They plan features when they could simply ask their agent for the result . He shows several concrete patterns:
- Audit your own agent history. A friend asked T3 Code for a feature that breaks down tool-call durations. Theo instead pasted the request into a thread ("break down all my tool calls… on this machine"), and the agent built a report page on its own. It found one home-directory search that ran for 41.7 minutes and another that ran almost 10 . Earlier audits showed that most of Claude's tool-call failures on his machines came from ripgrep commands written for bash that broke under zsh .
- Ask for a context-window visualizer. Doing this showed that, because of how his proxy was set up, Railway had been adding 70,000 tokens of context to every request for two months . Turn any request you keep repeating into a skill or a CLAUDE.md entry .
- Describe the delegation in the prompt instead of waiting for a subagent picker. His example: "spin up two opus 5.5 delegates to explore performance wins in T3 code, summarize their thoughts, and then start a new thread using Grock 4.7 to implement your findings" .
-
Write support flows as a markdown playbook.
npx t3 triagecollects basic machine facts and passes a prompt to Claude Code or Codex. The playbook tells the agent to ask what went wrong, get repro steps and screenshots, read the facts and the source, check whether the bug is already fixed upstream, strip secrets before filing, and comment on an existing issue rather than open a duplicate . A feature he expected to take "thousands of lines" became five lines of code and about 30 lines of markdown . - Ask the agent what it can do. He had the Codex desktop agent describe its computer-use capabilities, then pasted that description into Opus and asked it to "clone as much of this as you can". He says that work is still in a PR .
His larger point is to spend engineering time on plumbing that agents can use. He connected email to ChatGPT, has Dot watch for an AWS bill and investigate the costs, and connected T3 Code's MCP so Codex and Dot can start T3 Code cloud threads that ship changes .
Agent-scale throughput is breaking human-scale infrastructure
Anthropic went from about 2,500 merged changes a day in May to more than 30,000 a day in September. It says this left it bottlenecked by developer infrastructure built for humans, so it built its own version-control system and fully moved its monorepo off Git in five months (write-up) . Theo's reading is that the bottleneck was GitHub and that "a lot of us are gonna end up where they are today" . In a smaller example, Jullerino says his agent merged almost 400 web-performance PRs in 30 hours. He reports 86% fewer renders when another thread is streaming, up to 35% less JavaScript CPU and a heap about 10% smaller .
Geoffrey Huntley has a "spicy" fix for merge conflicts at that volume: throw away the active agent's working copy and make it rebuild and reapply its changes from the context it still holds in memory . Kent C. Dodds noted that, by Anthropic's own figures, 10% of its code is still written by hand: "What are you guys doing!?"
Defensive security loops with Daybreak
Huntley is running loops with OpenAI's Daybreak Blue to audit all of Overalls and says little is coming out. He adds that a clean result doesn't prove the code is secure . His operating assumption is that autonomous attackers will look at every commit of released source, possibly with more inference capacity than you have. So what matters is how fast you respond. He recommends running self-hosted tools on private Tailscale or WireGuard networks and doing autonomous blue-team exercises on your own code .
OpenAI's Tibo says OpenAI worked with OpenCode to get OpenCode users who are also OpenAI API customers working again after an attack, and to refund any credits the attack would have used. He urges teams to request Daybreak Blue access before they have an incident . Individuals can sign up at chatgpt.com/cyber, and organizations can apply through trusted access .
Overalls and inference routing
- Overalls (Huntley's self-hostable agent platform) can now pause and resume Firecracker VMs. Thread-created machines suspend after 4 hours with no tool calls. A new prompt, or an incoming VS Code web, SSH or web-app request, wakes them up. Machines can be forked along with their live memory state, and they move to another bare-metal node when capacity runs out . It can now also run Android as a Firecracker guest, displayed in the browser .
- geoclass early result: at 2 p.m. in Australia, capacity was short. Routing to Germany got around it, and he plans to switch back once Germany wakes up . His Underclass usage of OpenAI models over about 7 days came to roughly $39,314 at retail pricing .
T3 Code: performance demo and a disputed benchmark
In Theo's speed demo comparing T3 Code with Codex, the T3 Code threads ran on remote VPSs and Linux boxes, while the Codex side was mostly local. The two setups weren't the same . A benchmark fight is unresolved. David Ondrej alleges that Theo paid for a v2 run and asked for Cloudroom to be hidden after it beat T3 Code in v1 . Theo denies this. He says the evaluator had tested a T3 Code version more than three months old, and that he offered to fund a real benchmark . Treat both rankings with caution until a methodology is published. Theo also says that over the last two weeks, more than 60% of T3 Code users who used Codex also used Claude Code .
Smaller items
- Kody: Kent C. Dodds removed the inline UI generation built on MCP Apps, expecting agents to ship their own version. He is positioning Kody as "the shared cloud for your teams agents" .
-
llm-echo 0.5, Simon Willison's fake model that repeats prompts back without calling an LLM, now counts tool calls, system prompts and reasoning tokens in its simulated token usage. Try
uvx --with llm-echo llm -m echo 'hi there' -o thinking 1 --usage. llm-tools-datasette 0.1.1 fixes a break caused by LLM 0.33's switch to httpx2 . - pi-durable will get proper hibernation on serverless platforms such as Cloudflare Durable Objects and Vercel, plus telemetry, next week .
- In the desktop app, you can now edit earlier messages in a thread, per Tibo .
- One user plugged a Kindle in over USB and had Claude Code jailbreak it, strip out Amazon's bloat and add Tailscale .
- Ask for outputs rather than building dedicated features: Theo prompted T3 Code to turn machine tool-call history into a duration report; it surfaced one home-directory search that ran for 41.7 minutes and another that ran nearly 10 minutes. He also prompted a context-window visualizer and says a useful recurring request can become a skill.
- For performance work, he asked an orchestrator to launch two Opus 5.5 delegates to explore server and client improvements, summarize their findings, then start a new thread with Grock 4.7 to implement them; specify the desired delegation in the prompt rather than waiting for a dedicated subagent selector. Repeated instructions can be moved into skills or agent configuration.
- For support triage, T3 CLI collects basic machine facts and pipes a prompt into a chosen coding-agent harness; its Markdown playbook has the agent gather the issue description and repro steps, inspect machine facts and source, check upstream, and file an issue when appropriate—after removing secrets and checking for duplicates. Theo says this replaced a planned elaborate flow with five lines of code and about 30 lines of Markdown.
- Build access and data plumbing so agents can act across tools: Theo connected email to ChatGPT, used an agent to monitor for an expected AWS bill and investigate high costs in the AWS dashboard, and linked T3 Code through MCP so other agents could start coding threads and carry work through to changes.
MEGA's ad argues that when agents write faster than people can read, adding subagents or parallel sessions will not remove the human bottleneck; it proposes async, remote agents connected to apps, services, and other agents, supported by a managed knowledge base that keeps agents informed about the software. This is a program's promotional recommendation, not a report of a tested workflow.
llm-echo 0.5 is a debug model that repeats prompts without calling an LLM; its simulated token-usage counts now account for tool calls, system prompts, and reasoning tokens, making it useful for testing prompt and usage handling . Example invocation: uvx --with llm-echo llm -m echo 'hi there' -o thinking 1 --usage .
The ClaudeDevs account reports that Claude writes about 90% of Claude Code’s code; from June to September, merged PRs more than doubled while the share reverted fell by more than 40% .
llm-tools-datasette 0.1.1 fixes a compatibility break after LLM 0.33 switched to httpx2; the plugin exposes Datasette instances to LLM as a tool and had used https without depending on it directly.
- Geoffrey Huntley says he is running audit loops with Daybreak Blue across Overalls; so far, little has surfaced, but he cautions that a clean audit does not prove the code is secure. He contrasts the project’s four years of work by a team of at least 10 under pre-AI accounting with one person doing it in under a month now.
- His security takeaway is to prioritize response speed and assume autonomous attackers inspect every commit of released source code. He proposes autonomous blue-team exercises against source code using vendor-sponsored inference and ablated models, running Overalls on private Tailscale networks—or WireGuard/tinc—instead of the public internet.
- He says he is working with Anthropic to get Overalls into Glasswing and asks for support on oss-scanner PR #704; this is an effort in progress, not a completed integration.
Kent C. Dodds says he removed Kody’s MCP Apps-based inline UI generation because he expects coding agents to offer a first-party version; he positions Kody instead as a tool for gaps outside chat, especially helping teams use another agent through a shared cloud for their agents.
Overalls now supports running Android as a Firecracker guest on Linux, with software rendering and display and interaction in a browser. Huntley has started building a mobile client for iOS, Android, and desktop; the post does not specify the implementation stack.
- Geoffrey Huntley released GeoClass, an experimental prototype for routing requests across inference regions, as a way to pool OpenAI subscriptions and work around regional capacity limits. The setup uses a Tailscale network with tagged VMs deployed in multiple regions, then points Underclass at GeoClass .
- In a reported test, Australian inference capacity was short at 2 p.m.; switching to Germany resolved the capacity issue. Huntley planned to switch back to Australia when Germany woke up .
Theo says he investigated a low T3 Code benchmark result and found the evaluator had used a version more than three months old; the evaluator then tried the modern version with T3’s new orchestrator, a caveat for comparisons across agent versions and setups. Theo denies an allegation that he asked the evaluator to hide Cloudroom and says he offered to fund a real benchmark because good research was lacking.
Geoffrey Huntley shared a link whose title frames running your own LLM proxy as a way to understand what is happening—an observability-oriented LLM tooling signal, but the post gives no coding-agent workflow or implementation details.
pi-durable plans next week to add proper hibernation support on serverless platforms such as CF DO or Vercel, plus telemetry.
Overalls.dev now supports pausing and resuming Firecracker machines . In /threads (Orb) workflows, machines suspend after four hours without a tool call and resume on a new prompt/tool call or an incoming request such as VS Code web, SSH, or a shared web app . Machines can be cloned with live memory state and controlled via SSH; if a Lilly node lacks capacity, they can be relocated to one that has it, with the rules applying after relocation . The project presents this as self-hostable resource management to address the cost of CPU and memory .
Overalls.dev, described as an AI-focused self-hosted developer platform, announced internationalization support and listed flags for the US, China, India, Spain, Saudi Arabia, France, Bangladesh, Brazil, Portugal, Russia, Pakistan, Indonesia, Germany, Japan, Nigeria, Turkey, Vietnam, South Korea, the Philippines, Estonia, Finland, and Sweden.
Riley Brown reports editing a video with one prompt in Opus 5.5, starting from a single take of himself with many mistakes. The workflow analyzed the transcript, cut around mistakes and word gaps, and used Claude Code to search Twitter for relevant B-roll via GrokBot before scraping the video assets. It built a four-shot Blender scene from a template (computer, phone, table, and board), used screenshots to inspect and fix sections, and synchronized shots, zooms, and animations to the spoken words. It rendered 3,487 frames on RunPod cloud GPUs in under 20 minutes for less than $1, then added word-by-word captions, sound effects, and music and uploaded the result to Google Drive.
A Kindle mod used Claude Code to jailbreak a Kindle connected over USB, remove Amazon bloat, and add Tailscale—a concrete example of using a coding agent to customize an existing device. Riley frames this as part of a broader pattern of using AI to remove unwanted software from devices and replace apps with agents.
For merge conflicts, Geoffrey Huntley proposes clobbering the active agent’s working copy, then having the agent use its in-memory context to rebuild and reapply its changes .
- Theo says T3 Code’s latest nightly felt “insane” fast, but gives no benchmark; the T3 Code threads shown in his demo were running on other machines, mostly VPSs and local Linux boxes.
- His Codex comparison was mostly local, with a few threads on an overspecified Windows machine connected directly over 10GbE, so the setups differed.
Theo says the evaluator’s low T3 Code result used a version more than three months old; after trying a newer version with T3 Code’s new orchestrator, the evaluator found the scores were still unfavorable and adjusted. Theo says he offered to fund a proper benchmark because he sees a lack of good research, and that the evaluator surfaced issues he had been investigating.
David alleges Theo paid for a second benchmark after Cloudroom ranked above T3 Code in the first, and asked for Cloudroom to be hidden. Theo denies asking for Cloudroom’s removal, says he did not know what it was, and says his only comment was surprise that a clone of T3 Code initially scored higher. This leaves a material but disputed benchmark-integrity concern; the posts provide no benchmark results or reproducible methodology.
Riley Brown says his best AI-agent workflows involve sensitive data, so he cannot show them in content; he does not describe the workflows themselves.