We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Models are routing around the edit tool
In State of Agentic Coding #11, Armin Ronacher reported that Pi still provides an edit tool, but with Astra and Opus 5.5 it goes unused in about 80% of cases. The models write one-off Python to edit files instead . His explanation is that training rewards success by whatever means, so the models fall back on ad hoc code. Pi already offers "code mode", which runs JavaScript in the harness loop rather than bash where your files live, and will probably make it the default .
Three consequences for developers follow:
- Readability. Astra's code written on the fly reads like code golf: single-character variables and no indentation . Ben Tossell pointed out the tension: very compact code suits a tool loop but isn't what you want in application code . Watch for it leaking into your codebase.
- Visibility. Pi will probably hide tool output soon, because when the model runs unreadable code, explaining it would take a second LLM . Ronacher calls this a regression. For critical work you need checks along the way, not just the agent's word at the end, and he floats a separate commentary stream from another model .
- Model divergence. Huntley says Kimi still reaches for the edit-file tool, while the frontier American models write code over bash to get around it .
Huntley's write-up adds a hot take you can act on: delete and redo your AGENTS.md and many skills after every model release, or at least every couple of months, because "best practice" goes stale . The same reasoning applies to harnesses. Most agents copied Claude Code's loop: transcript, tool call, tool result. That loop fits background compaction and multi-person Slack threads poorly, and an agent will "happily build you a decent-looking agent" on top of it that is unreliable underneath . Pi Durable, which resumes exactly after a harness restart, including with sub-agents talking to each other, took four months for about 10,000 lines. The team solved the problem around twenty times, about two hours each, before the trade-offs felt right .
Two other ideas from the episode:
- Huntley's Lisp demo connects an LLM to a running Lisp image with just two tools, evaluate an expression or define a function, and lets you roll changes back. You change the running program by chatting with it instead of editing files and recompiling .
- Ronacher uses fast classifier models each morning to check open issues against his own criteria and see how they changed since the day before .
Overnight and parallel patterns from High Taste Testers
On Latent Space's new High Taste Testers show, one practitioner said Opus 5.5 is finally good enough to start agents at 10 p.m. and review at 9 a.m. across about 10 prototypes. He also warned that agents running non-stop on production software create more code than you can absorb . His method is a "poor man's Ralph loop": queue 5–10 prompts, such as implement, compact, reread the plan, keep improving, so the run stays bounded .
Other patterns from the episode:
- A "detach mode" skill runs CLI agents in the background, up to 40 in parallel, with logs to monitor, pause, improve and rerun .
- Codex staff reportedly said subagents beat separate sessions because of cache optimizations . One user set 48 subagents in their config and burned through their Astra limits in about two hours; another keeps it at about 15 .
- For very hard problems, run the main thread in ultra and give it only one job, dispatching subagents. Put timers on them and push any that stall .
OpenAI's Tibo says Codex & Work now gives more effective usage by reusing context across subagents, keeping context cached longer, and routing requests to get more cache hits . ChatGPT's GPT-6 Astra Pro can now run several agents in parallel on parts of one task, for Pro users in the Chat tab .
Review and codebase hygiene
The Claude Code team (Addy Osmani, Morgan Westlee Lunt and Qing Wang) says "did a human read the code?" is the wrong question. The real one is what could go wrong with a change and which check, automated or human, should catch it. Review now runs continuously "like spellcheck", and Claude has reviewed and fixed most changes at least once before a human looks . Boris Cherny says Claude Code Review runs on all his PRs and catches almost every bug before it goes in .
Geoffrey Huntley points out that agents copy a messy codebase's bad patterns everywhere. He suggests periodically having the model rename things and restructure modules, files and directories. He calls this a "janitorial Roomba" job, suited to a skill run by cron or by a GitHub Actions job on trunk . His personal-bot write-up has one design rule: build thin bots, one per function, that call each other and chain together like Unix pipes . His Overalls bot sends coding tasks to short-lived Firecracker VMs on his own infrastructure . He also flags Claude Code's crossSessionInbound=accept setting as apparently the key primitive behind "hive swarms" .
Tool releases
-
Cursor: cloud agents can build interactive canvases, such as dashboards or reports, that you share by link . The Environments API (beta) lets you create and manage environments programmatically, including the repos they clone and their
environment.jsonsetup . Environments now support up to 500 repos, and large multi-repo environments start 5x faster (changelog) . - LangSmith AI Search: structured filters narrow the traces first, then a small model checks each remaining one in parallel, usually in under half a second each . LangChain checked 150 conversations from its Open SWE coding agent. The median ran 84K tokens, about 1.5K came from the user and the final reply, and roughly three-quarters was tool output .
- Managed Deep Agents adds your credentials at a proxy outside the sandbox, and only on calls to services you've allowed, so a malicious package can't steal them .
- Riley Brown's hands-free setup: walk with AirPods and talk to his Dot coordinator, which starts several Codex sessions, such as company research, a landing page and a website update. He says to connect your tools so it has context . He also calls GrokBot through a webhook skill from Claude Code to search X without an X API key, and says setup takes under two minutes .
- Fai (repo) is Krzysztof Cieslak's functional language for exploring software engineering tools designed for agents .
- Theo says OpenAI has lately given him significantly more cybersecurity risk flags and refusals than Anthropic .
- Pi's team spent four months and about 10,000 lines of code building a durable agent architecture; although agents could produce many candidate solutions quickly, settling on the right trade-offs took much longer. The framework lets the agent resume after a harness restart, work over a network, and support communicating subagents; it was not yet integrated into Pi's coding agent because it would break several extensions.
- Pi's original transcript-and-tool-call loop, modeled on Claude Code, fit poorly with background compaction, multiple people using one agent, and branching work involving subagents. The discussion recommends revisiting foundational harness assumptions and rethinking agent instructions and skills as models change, rather than treating old practices as permanent.
- Pi offers code mode as an option and is considering making it the default: it runs JavaScript in the harness loop rather than Bash against the project filesystem. The discussion reports that models increasingly synthesize code for file edits instead of using supplied edit tools; this can make their actions difficult to inspect, and hiding the raw tool output is described as a visibility regression.
- A practical issue-triage workflow uses fast classifier models to review open issues against custom criteria and check how they changed since the previous day.
- One practitioner found Opus 5.5 good enough for overnight work: while developing about 10 prototypes, he planned to start agents at 10 p.m. and check their work at 9 a.m. His bounded “poor man’s Ralph loop” queues 5–10 prompts—such as implementing, revising, rereading the plan, and iterating—rather than running an open-ended loop; he cautions that agents can generate more production code than he can reliably consume.
- For parallel work, one practitioner detaches CLI agents, runs up to 40 in parallel, monitors their logs, and pauses, revises, and restarts them; another pattern is to keep the main agent focused on dispatching and reviewing subagents, with timers and follow-up prompts for stalled workers. A 48-subagent Astra configuration reportedly exhausted its limits in two hours; a participant said they had reduced the count and another kept it around 15.
- For structured classification and model switching, use existing labeled examples as an eval: compare a cheaper candidate against historical labels, automate when agreement clears a chosen threshold, review disagreements, and add failures as eval cases for prompt iteration. A play-tested agent-memory pipeline used a decision model for structured extraction, sorting, and reranking alongside Soul 5.5/5.6 or Soul 6; its builders reported 10–20× lower cost and alignment with Opus 5.5 in their evals without reasoning.
- Huntley reports that Ralph loops can build unattended overnight; Dex estimated a run at about $10.42 per hour.
- Pi Durable is designed to resume exactly after its harness is shut down and restarted, including across network use and complex sub-agent interactions. The team spent four months on roughly 10,000 lines of code because finding the right API and trade-offs took repeated experiments, even though each attempted solution came together quickly.
- A basic transcript/tool-call/tool-result loop handles newer demands poorly, including background compaction and multi-person Slack sessions with branching conversations; Huntley argues that agent builders should rethink that underlying primitive rather than keep copying it.
-
Huntley recommends revisiting or redoing
AGENTS.mdand skills after model releases—or at least every couple of months—because agent best practices and assumptions can become stale quickly. - In Huntley’s live Lisp demo, the LLM edits a running Lisp image using just two tool calls—evaluate an expression or define a function—and changes can be rolled back. This enables evolving a running product through chat instead of the usual file-edit, compile, and review workflow.
- Pi’s edit tool is reportedly unused in about 80% of cases with the latest models; models instead synthesize code to make changes. Pi may make JavaScript code mode the default in place of bash, but compact code that works well in a tool loop can be unreadable and unsuitable for application code.
- For critical agent work, verify progress along the way rather than relying only on the final result; as agents expose less tool output, one proposed approach is a separate commentary stream explaining the generated code, though that may cost extra tokens.
- A voice-first delegation workflow: walk with AirPods and talk to a coordinator in one thread, asking it to launch multiple Codex sessions in sequence. The speaker’s example splits work into researching a company, creating a landing page, and updating a website; he says the agent can access his locally running computer and recommends connecting tools so it has context.
- For coding, the speaker prefers Opus 5.5 in Claude Desktop and recommends Claude Code or Claude Code Projects for complex work; Projects provides a coordinator that can spin up multiple threads. He reports moving faster and getting more done with the Claude Desktop/Opus 5.5 setup.
- A coding agent can use a GrokBot webhook to search X: the speaker demonstrates invoking the webhook skill from Claude Code and says the setup takes under two minutes, works with Claude Code or Codex, and avoids needing an X API key or extra search fees.
- In a “hive coding” example, creator Max Blade gathered audience suggestions for his Kong Royale battle-royale game during a livestream, selected the best requests, and let them run overnight; the speaker says the first pass made 60 changes.
Huntley’s upcoming Overalls (formerly loom) is a remote interface for dispatching agentic coding tasks to ephemeral Firecracker VMs on his self-hosted infrastructure. His reusable design rule is to keep bots thin and function-specific, then have them call one another in Unix-style chains; he also uses Tailscale on his Grok bot for tunneling. He avoids self-hosting OpenClaw while the “lethal trifecta” security problem remains unresolved, and says he would move to self-hosting only if it is fundamentally solved.
- Riley recommends Codex or Claude Code; for complex work, he favors Claude Code Projects, where a coordinator can spin up multiple threads. He says Claude Desktop running Opus 5.5 is helping him move faster and get more done.
- His hands-free workflow is to connect the voice coordinator Dot to his tools and computer, then talk to it through AirPods while walking. From one thread, he can start multiple Codex sessions—for example, researching a company, creating a landing page, and updating a website.
- A cross-agent research setup uses a GrokBot webhook skill from Claude Code to search X without an X API key, account connection, or extra payment; the creator says it works with agents such as Codex and can be set up in under two minutes.
- Max Blade’s “hive coding” project used a livestream audience to suggest changes to a 20-person battle royale game; he selected suggestions, let the agent work overnight, and the first run made 60 changes.
- Anthropic’s Claude Managed Agents dynamic workflows became available in public beta. The orchestration pattern uses a lead agent to plan work across multiple agents in phases, then combine their results.
- Swyx promoted a full-length Claude Flows workshop by CJ and Harrison Stall at AI Engineer NYC; a follow-up linked a YouTube video.
Swyx announced a High Taste Testers episode in which AI power users compare what works in practice; the listed guests include Dex Horthy, and the agenda covers coding tools, subagents and context management, agent loops and forks, local-model coding benchmarks, and delegation and computer-use workflows .
The Claude Code team describes shifting review from an end-of-PR human gate to continuous review: Claude reviews and fixes most changes at least once before a human looks, catching routine mistakes and leaving human attention for higher-level questions. They frame review around what could go wrong and which automated or human check is best suited to catch it, rather than whether a human read the code.
A user linked a leaked key and a $90,000 Google bill to Opus 5.5’s claim that Codex had published their .env to Cloudflare Pages; the user said Codex could not find the issue, refused to explain, and flagged the chat for cybersecurity risk. ThePrimeagen echoed the exposure claim, but the posts do not independently verify it.
After spending a day on agent setup, Tossell says it is paying off. @bot opens threads in @get_bb_app for “proper” coding work; the thread emails bot updates and updates a shared task list and work log.
Geoffrey Huntley is the guest on State of Agentic Coding episode #11; the episode announcement lists code mode and self-modifying harnesses among its topics . The post’s teaser also names “the curiosity test, lisp, and labs that can't fail” .
Krzysztof Cieslak introduced Fai, a functional-programming-language side project exploring how to design better “classical” software-engineering tools specifically for AI agents; the project is at https://github.com/Krzysztof-Cieslak/fai. Geoffrey Huntley identifies Cieslak as the original researcher and creator of spec-driven development.
Geoffrey Huntley linked a firsthand “how I use @bot” write-up and disclosed that @poteto supplied enough credits for him to revisit his priors about personal assistants’ utility .
ThePrimeagen says he includes the same picture whenever he talks to an LLM, using it to set ground rules before proceeding; the post does not provide the rules in its text.
- For continual agent training, SFT/distillation can teach behaviors from gold traces the target model would not generate itself, but its off-policy data may leave the model unable to recover from its own mistakes. RL learns from the model’s own sampled rollouts and is reported to cause less forgetting, but it cannot reliably teach behaviors outside the model’s existing sampling capabilities or bake in specific facts .
- In experiments described, distillation warm-started RL at a higher level, but the distilled model plateaued below the model trained with RL alone; the result depended on the task, and RL curricula may need to progress from tasks with nonzero pass rates to harder ones .
- The talk cites Cursor’s Composer as using on-policy self-distillation hints alongside RL rewards during rollouts, while warning that self-distillation can bias RL and be difficult to make work . For broad, tool-using agents, the speaker says teams may need custom RL environments and mid-training, then tune the SFT/distillation-to-RL mix to their task .
Bluesky promoted Attie as a social vibe-coding tool for making sites such as games, quizzes, guestbooks, and “top 8” pages. Geoffrey Huntley characterized Bluesky’s embrace of vibe-coding as contrary to its community’s wishes.
Geoffrey Huntley reports that a “brilliant intern” autonomously fuzzed and beat Antithesis’s “golden goose” game, which most people had considered too large in state space to beat autonomously.
A poorly structured codebase can cause agents to replicate bad patterns; periodically use an agent to improve names, module and file organization, and directory structure. Huntley calls this a “janitorial Roomba” task and suggests packaging it as a skill run on a background cron job or as a GitHub Actions job off the trunk branch.
Boris Cherny says he now makes slide decks with Claude Slides rather than googling for assets such as an iPhone-frame SVG; he points users to Claude Artifacts.