We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Keep a read-only "explainer" chat next to the agent that's building
Salvatore Sanfilippo (antirez) describes a problem many agent users run into. After the implementation agent finishes a step, the best next move is a follow-up prompt while its context is fresh. So you skip asking it to explain what it did, and gaps in your own understanding of the project pile up .
His fix is a second chat on the same project, opened read-only. You tell it that someone else is working in the repo, that it must not write or touch anything, and that its job is to explain architecture and how each part is implemented . You ask the same things every time, so he says this is one of the few cases where a skill file is worth writing. His skill covers:
- never edit the project;
- if details don't match, assume the other agent made an edit and re-check;
- no jargon; aim to build a mental model in the person asking;
- explain specialist details in a short aside;
- don't re-explain topics you say you already know unless asked .
This chat can run on a cheaper model. A weaker model may not manage a hard implementation, but it can usually explain one that already exists, for example Sol instead of Astra when you're near a weekly limit . He doesn't open it during the first hours of fast iteration. He opens it once long-horizon features start, and says it changed how well he understands details he used to leave to Opus or Astra . He pairs it with a Claude hook that plays an audio message when Claude is waiting for input .
In a separate C lesson he applies the same approach to new code. He works through the low-level details by hand until the architecture is clear, then uses a coding agent to move faster .
Effort levels: practitioners are moving off the maximum
After Theo's switch from extra-high to high effort on Opus 5.5, swyx went further: "low effort + critique loop > superultramaxhighwhatever." He frames this as advice for throughput-constrained users and points to his swyxio/skills/babysit-runs skill . He also reports that Karpathy used medium effort on Opus 5.5 for Hackertube, launched that day. His overall advice: "just tweak to taste" .
Arca's production harness: patterns worth borrowing
On Latent Space, Aman Singh described how Arca runs agents on financial work that spans days. Several patterns carry over to coding-agent infrastructure:
- A model-agnostic harness with routing by task tier. Heavy analysis gets the top tier (Opus-, Sonnet- or heavy-GPT-class models) and background work gets lower tiers . Internal benchmarks drive model choice. Luna looked strong on external benchmarks but was too slow and expensive in their use .
- Shared memory for long-running work. For each client they store what's happening, who owns what, and what comes next. A nightly "dreaming" process updates it, so a newly started agent can catch up quickly .
- Guardrails at the edges. Deterministic restrictions decide what agents can do. Fast LLM checks review every database write. Incoming content is screened for prompt injection, and outgoing data gets PII checks plus a consensus of small models looking for exfiltration . Code the agent writes in the sandbox is translated into a strict format and then discarded before it reaches the backend .
- Traces become improvements. Staff feedback is stored alongside the trace that produced the work. A meta-agent runs over each night's traces and suggests harness or knowledge-base fixes. A human engineer still reviews prompt changes, because agent-written fixes tend to bloat the prompt: "every little fix becomes the whole prompt" .
Tool releases
- Kody v2026.10.31 ships profiles. A profile is an isolated identity an agent acts as, with its own grants, expiry and revocation, and profiles can chain into sub-profiles (release) .
- OpenCode now accepts Claude subscriptions . Theo reads this as Anthropic "loosening the grip on the sub a bit" .
-
T3 Code added
/btwand side chats, which work with any model or harness . This is a built-in way to ask questions on the side, close to Sanfilippo's pattern. Theo also shipped a fix (PR #28409) after a user showed the model picker breaking on providers with very long model names and option lists . - ChatGPT Dot → Codex: Riley Brown sends Dot an iMessage voice note asking it to start a local Codex session, tidy his landing page and push the changes . He says Claude Duet can't yet start local Claude Code sessions. Anthropic's Gabe told him remote control and local Mac access are in progress .
- Hex, a fast, local voice-dictation tool for talking to agents on Mac, Windows and Linux, has a stable release . DHH added it as a dictation option in Omarchy Dev (PR #17186) .
Model takes
- GPT-6.1 Sol vs GPT-6 Astra: Will Depue says Sol is "mostly as smart," faster and easier to understand . Theo agrees: it "doesn't demo as well, but it gets to the right answer more often in real work" .
- Google Argon 4: ThePrimeagen praised it . Theo says it can't work in a real codebase without "egregious" hallucinations and is "10x worse," not 5% worse .
- Jerry Tworek says Core Automation has used Codex but is gradually moving to in-house tooling. His view is that Codex is a general tool that "doesn't add that much value on top of just the model" .
Short notes
- ThePrimeagen started "agent directed walks" for Omarchy QA. He expects it to be the largest and hardest part of QA to get right, and plans to post long runs early in the week .
- Geoffrey Huntley cut overalls.dev builds from 28 minutes to 78 seconds with Buck2/RBE. He is now looking at a ~266 GB closure size . Fast builds matter more when agents rebuild constantly.
-
Security: Parameter Labs broke out of ChatGPT's sandboxed HTML previews. A message handler in the MathLive library used
innerHTML, which let them inject an unsandboxed iframe, and a YouTube oEmbed endpoint got around the CSP . A crafted share URL could steal a signed-in user's data or act as them. OpenAI fixed it and paid a $5,000 bounty . If you show agent-generated HTML in a sandboxed iframe, this attack path is worth reviewing.
- Arca’s in-house harness is model-agnostic and routes by task complexity: tax analysis uses the heaviest model tier, while background tasks use lower tiers. Internal benchmarks guide model choices; one model with strong external benchmarks did not meet expectations on speed and cost.
- For work that runs across days, Arca persists each client’s current situation, owners, history, and next actions, then uses a daily reflection process to update long-horizon memory. Responses in an existing email thread go to its agent; a fresh signal, such as a phone call, can wake a new agent that consults the shared state and connects the signal to ongoing work.
- Arca keeps money movement, asset liquidation, and direct client communication human-gated while agents handle analysis. Its safeguards include deterministic restrictions, verification of database writes, inbound prompt-injection checks, and outbound checks for PII or attempted data exfiltration. For sandboxed analysis, agent-written code is converted into a strict expected format and discarded before the backend receives it; generated artifacts are checked deterministically and with smaller models before publication.
- Arca pairs staff feedback with the trace that produced the work, then runs a nightly meta-agent over real-world traces to suggest harness, tool, or knowledge-base improvements; a human engineer reviews prompt changes before release. Experts curate operational knowledge in an editable product knowledge base, which agents cite and trace analysis can flag for correction.
- Riley Brown used ChatGPT’s Dot over iMessage, including voice input, to launch a local Codex session for minor landing-page design cleanups and push the changes; Dot could use connections he had configured in desktop ChatGPT.
- Brown used Claude Duet to plan a presentation, but at the time it could not launch local Claude Code sessions. An Anthropic representative said remote control and local Mac access were in development; Brown said he would switch from Claude Projects to Duet if local Code access arrived. He also liked Duet’s channel-based organization instead of using separate agents.
- For large projects, keep a second, read-only agent chat open on the same project while the implementation agent works. Use it to ask how features are implemented and to request brief architectural explanations; a reusable skill can tell it not to edit files, avoid jargon, explain specialist details only when useful, and account for what you already know.
- This separates learning questions from the implementation chat, preserving that chat for the next step and helping you stay in the project’s flow. A lighter, cheaper model can often explain existing implementation adequately, even if it may be less capable of implementing a complex feature. The speaker now opens this second chat on large projects when work moves from fast iteration to longer-horizon features, and says it has improved his understanding of details previously delegated to agents.
- Jerry Tworek said Core Automation had used Codex internally for a while and was gradually moving toward its own tools, which he described as more flexible and better tailored to how the team works; he considered Codex broadly useful but said it added limited value beyond the underlying model.
- Tworek said agents and automation help a small team build frontier systems, and that systems and data staff aim to use automation to do work equivalent to a team elsewhere. He also said modern AI solutions could potentially compress two years of work to three months when starting an AI lab.
ChatGPT’s interactive HTML previews had a sandbox escape: Parameter Labs found that a sandboxed preview could message a MathLive handler that used innerHTML, enabling injection of an unsandboxed iframe; a YouTube oEmbed endpoint then provided a route to JavaScript execution despite the site’s CSP restrictions. A crafted share URL could let the attacker access private user data or act on a signed-in user’s behalf; OpenAI closed the flaw and paid a $5,000 bounty.
- For C networking work, Sanfilippo’s sequence is to first work through the low-level Unix socket/UDP mechanics and settle the architecture, then use a coding agent to speed implementation; he says LLMs are unnecessary for learning the fundamentals, but understanding them helps with later decisions.
-
Encapsulate repetitive, awkward socket setup and
sendtodetails in a small reusable API: he builds a UDP helper/library and says the resulting C file can be reused across projects. He considers giving an LLM cleaner building blocks optional—the model can handle low-level code, while a mental model lets developers understand what it did without needing to inspect every system call.
swyx reports that Karpathy used medium effort with Opus 5.5 on Hackertube, which launched that day. For throughput-constrained coding-agent work, swyx favors a low-effort generation-and-critique loop over maximum-effort runs and points to swyxio/skills/babysit-runs; he recommends tuning effort preferences to taste.
Addy Osmani shared a Halloween post featuring Opus and three.js ; a reply links to the playable “Trick or Clawd” project: https://trick-or-clawd.claudepages.dev .
Anthropic rewrote its Messages API in Rust despite having no Rust developers: Claude wrote the code, humans wrote the plan and checked that the old and new versions matched on real traffic, and CPU use fell 40–50%. The account presents this as a counterexample to blanket advice against large rewrites.
-
Jason Zhou shared a Treg
rea-growthskill workflow for coding agents; the tutorial lists Claude Code, Codex, Grok Bot, and Hermes, and prompts the agent to set up Treg and install the skill. For a teardown, specify the company, geography and time period, set a spending cap, and require a price quote before each new data source. - To compare competitors, run setup once, then the company prompt for each of three competitors; review dated tactics marked “seen” or “inferred,” prioritize those repeated across all three, and skip one-offs. The tutorial reports about $0.50 per teardown and results in about five minutes.
ThePrimeagen briefly praised Google’s Argon 4, but gave no details about what it is or how it relates to coding-agent use.
DHH argues that refusing to comply with a company’s AI strategy should be grounds for firing, rejecting what he characterizes as a promise against job losses for AI obstinance . Proton’s assurance was narrower: it said it was not re-interviewing staff about AI and that there were no job losses due to AI re-interviews .
A T3code user called an issue their first complaint with the product . Theo said the video made it quick to fix and promised a fix by the end of the day , then followed up with “Done” and a link to PR #28409 .
Kody released profiles in v2026.10.31: an agent can act through an isolated identity with scoped access and full attribution, with its own grants, expiry, and revocation; profiles can also chain into sub-profiles. Kent described the feature as creating a token/MCP connection with only a subset of the user account’s permissions, to make agent access safer and more auditable.
swyx recommends using low-effort agent runs with a critique loop rather than maximum-effort settings, especially when throughput is constrained; he points to the babysit-runs skill for that use case. In the linked example, Theo says he switched Opus 5.5 from extra-high to high effort on Julius’s advice and found high better almost all the time.
Factory acquired Typedef, whose context layer helps agents understand enterprise data. Factory says the combination will give its Droids and engineers clearer lineage and evidence for planning migrations, debugging pipelines, and refactoring models; the post describes an intended capability, not a release or availability date.
ThePrimeagen began Omarchy QA’s “agent directed walks,” which he expects to be its largest and likely hardest-to-get-right part . He planned to share long runs on Monday or Tuesday showing features working .
Theo said he had not accounted for provider menus with extremely long model names or multiple sets of options in T3code, responding to a user's complaint about the UI. He called the user's feedback video helpful and said he would fix the issues ASAP.
Geoffrey Huntley says Buck2/RBE techniques remain “king”; builds for Overalls fell from 28 minutes to 78 seconds, and he sees room for further improvement. He is investigating closure size because a full build is about 266 GB.
- Riley Brown demonstrates using OpenAI Dot from iMessage—including voice messages—to start a local Codex session for a concrete coding task: tidy a landing page, make minor design tweaks, and push the changes. Dot can also use connectors configured in the ChatGPT desktop app, and send scheduled recaps or alerts through text messaging.
- A key limitation for Claude Duet is that it did not yet support starting local Claude Code sessions; after Brown raised this, Anthropic said remote control and local Mac access were in development. Brown considers local coding-session access essential to an agent platform. Duet’s channels let him organize work without creating separate agents.