We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Check your GPT-6.2 harness for a caching bug
If you run GPT-6.2 Astra or Sol through the Responses API, the top-level instructions field is not being cached. Codex issue #62312 found it stays at 0 cached tokens on every turn, even though OpenAI's diagnostics wrongly report a cache hit . The reported cost is large. Moving from GPT-6.1 Sol to 6.2 Sol dropped actual cache hit rates from 97.4% to 46.4%, so 64K–72K system tokens were billed at full price on every step .
Workaround: move the system prompt into a leading developer message in input. OpenCode already does this . Peter Steinberger landed a fix in OpenClaw's embedded harness and warns that "most harnesses aren't updated yet" . If you maintain your own harness, look at cached-token counts, not the diagnostics.
Gemini 4 Neon: hands-on reports
Last brief covered Neon's launch. This period brought firsthand usage. Matthew Berman built his open-source AI video editor, Scarlet, with Neon. He says the other frontier models keep needing "explain it more simply" follow-ups, while Neon both writes concise summaries and copes with "highly underspecified prompts" . His examples:
- Screenshot plus one line. He dropped in a screenshot of the model picker and asked for "make these one line and put the disabled ones below the active ones." About 5 minutes later the change was done. When he asked for a mock, Neon built a live one with toggles, favorites and filters, and said it would include a live mock for UI changes from then on .
- Filling in his intent. Asked to "add Gemini models to Scarlet like Anthropic and OpenAI are," it spent 45 minutes and routed editing through the Antigravity CLI over agent client protocol, plus the Gemini API for spoken-note cleanup. Berman says that was exactly what he wanted .
- Computer use. It diagnosed an OBS profile mismatch, fixed the settings, and did not restart recordings that were already running so it wouldn't interrupt them .
Berman says Neon is weaker on 3D demos and games . Riley Brown rates it "very good" for coding and video editing, but its "It's not this — it's this" habit is the worst he has seen, and he advises against using it for writing . That conflicts with Theo's setup in the last brief, where Neon wrote all the copy. Theo now reports "insane results" on long-horizon engineering work but hasn't shared details yet . He says Google is "so far ahead" on making the model easy to work with .
Clay: harness and interface work beat post-training
Clay's Jeff Barg, on LangChain's Max Agency podcast, gave the most reusable engineering lessons of the period:
- Post-training gave no step change. Their trained model moved closer to Sonnet/Opus, but the team decided "there's a lot of juice" in harness engineering instead .
- Make your domain look like what models were trained on. Filter-composition structured output handled easy searches but missed long-tail queries. Clay replaced it with a SQL-like query language, which models generate well and which is also composable and modular .
- Start small. They began with off-the-shelf Pi, chosen because it is light and extensible. They haven't forked it and customize mostly through plugins . The agent has about 20 skills and 8–10 tools, and 85%+ of calls go through bash . The production agent uses the same CLI that external agents get .
- Debugging long traces. Asking Codex what went wrong in a raw trace with hundreds of tool calls worked "really bad." Having it first summarize each bash command worked much better .
- Evals: Clay started with capability buckets, then tried coverage buckets per launch milestone, which Barg found "almost meaningless," and settled on user journeys split into subtasks .
Astra and Sol: what they're good at
Berman credits Astra's computer use with fixing frame drops he had chased for 18 months. It read OBS logs during a live recording and forced on a buffer setting that "Auto" had left off . Astra has powered Scarlet's video editing for weeks , although he rechecks Astra-built code with Opus . In a same-prompt comparison he shared from another user:
| Model | Time | Cost |
|---|---|---|
| Astra | 45 min | $9 |
| Sol | 30 min | $2 |
| Fable 5.2 | 3 h | $200 |
| Opus 5.5 | almost 3 h | $31 |
Berman judged Fable's output best, with Astra comparable .
Tools and releases
- LangSmith Tool Gateway (public beta) is a managed layer that handles credentials, identity and access for every agent tool call . Setup: open Tools, add an MCP server or turn on a built-in, connect your accounts, pick the tools for a kit, and point any MCP client (including Claude Code or Cursor) at the kit's URL . Search mode loads only the tools relevant to the current task . Admins can deny tools or require human approval for sensitive calls .
- Antigravity desktop can now connect to remote servers over SSH .
-
Solid 2.0 errors include an error code, the exact location, a fix, and a path to a
skill.mdthat tells your agent how to fix it. Some of the last scheduler bug reports came from agents . - JetBrains Air (from a sponsor segment) runs Claude Code, Codex, Copilot and other agents side by side inside JetBrains IDEs, with access to tests, inspections and the debugger. Line-level review notes go back to the agent .
-
treg
/doomscrollis an open-source skill: give it a topic and a time window, and it finds posts across X, TikTok, YouTube, Reddit, HN and Product Hunt . Repo: superdesigndev/treg .
Security and quality signals
The Pragmatic Engineer covered the Railway hack. An attacker using an abliterated model (one with its safeguards removed) planned a route to AI provider keys within six hours of downloading Railway's monorepo . Railway's fixes:
- each build gets its own throwaway VM
- fine-grained access control for CI secrets, with 2FA or biometrics for sensitive credentials
- security designed on the assumption that attackers already have the source code
Greg Brockman posted "Codex for securing Ruby," quoting a researcher's universal pre-auth Rails RCE. Neither post explains how Codex was used .
Sourcegraph's 2026 Big Code Report says 54% of teams using AI find that more than 60% of AI-generated work needs significant rework, and only 10% say their agents share a common view of the codebase .
- Berman used Astra’s computer-control capabilities to investigate 18 months of recurring frame drops: it checked camera settings and the HDMI feed, then read OBS logs during recording. It found the OBS buffer disabled despite being set to Auto and forced it on, resolving the issue after hardware replacements and human experts had failed to identify the cause.
- Berman’s open-source AI video-editing app Scarlet used Astra for weeks; he said it cut dialogue well and was close to editing his videos end to end in one pass, though the app was still early and rough.
- In one same-prompt comparison, Astra took 45 minutes and $9, Soul 30 minutes and $2, Fable 5.2 three hours and $200, and Opus 5.5 nearly three hours and $31. Berman judged Fable visually best and Astra comparable, suggesting a speed-and-cost tradeoff rather than an unqualified quality win.
- Linear’s Loops can automatically assign an agent to an issue and start it coding; Triage Intelligence can label incoming bug reports and feature requests, assign them, and route them to the right team queue without manual triage.
- Linear said customers wanted APIs to build their own agents into Linear while retaining their proprietary data; it built an agent API and partnerships bringing coding agents such as Cursor and GitHub Copilot into Linear.
- Linear reported that agents created more than half of its issues, while app usage also increased; it tracks MCP and app usage together because teams still need the app to visualize and collaborate on work that a chat interface handles poorly.
- For coding sessions, users can choose OpenAI or Anthropic models and harnesses; Linear says it charges based on the API cost and published rate, plus a small markup. Its COO also said Linear’s purpose-built workflows make its data easier for agents to interpret than highly customized systems with inconsistent meanings for labels.
- Theo reports that Claude writes more than 90% of his code. When the agent gets something wrong, he says he either directs it to correct the issue or writes that part himself.
- Theo endorses treating prompting as a search for inputs that produce the desired code without hand-editing; he compares it to speedrunning and argues that automating more work creates room to experiment with more ideas.
- In a sponsored segment, Theo identifies GitHub rate limits and slow runners as bottlenecks, then pitches Depot: Docker-layer caching for development, CI, and agents; DepotCI, which he describes as faster and easier to parallelize than Actions; and planned Depot Code, a GitHub-API mirror intended to work with Depot CI and sandboxes for immediate code access.
- While building Scarlet, an AI video-editing app, the presenter says Neon handled underspecified coding requests: a screenshot plus a short layout instruction led to updated model settings and a live mock with toggles, favorites, filters, and provider switching.
- He asked it to add Gemini models alongside Anthropic and OpenAI; after about 45 minutes, it routed editing tasks through the Antigravity CLI and an agent client protocol, and spoken-note cleanup through the Gemini API—choices he says matched his intent.
- In an OBS computer-use task, he asked the agent to inspect a sample recording and then fix the settings; it identified a profile mismatch, operated OBS, and avoided touching or restarting the active recording windows to prevent interruption.
He keeps SSH closed by default, opening it only after an undisclosed, counterintuitive action; he says he built this mechanism to his own specifications and audited it with AI rather than adopting a tool.
As a proposed workflow, he would have frontier LLM agents check a server hourly for signs of attacks, try to block threats, and send Telegram alerts; he describes giving the agent a Telegram bot API key and his Telegram user ID for notifications. He also suggests asking an LLM to rewrite part of a parser so exploits no longer match changed offsets or triggers.
Distinctive prompts can serve as model fingerprints: ChatPRD’s agent received requests for SVGs of pelicans riding bicycles, and Willison says he has used similar probing to guess the identities of stealth-preview models on OpenRouter; the suggestion that token abusers used this specific prompt to identify models is speculative.
The quoted post argues that Three.js demo ratings are a poor basis for judging models and that RL-developed Three.js capabilities do not extend to real software . ThePrimeagen replies, “What if the demo looks like GTA6 though?”—a challenge to dismissing demos, but not evidence of real-world coding ability .
-
Solid 2.0 is designed for agent-assisted maintenance: its errors provide an error code, exact location, suggested fix, and a
skill.mdpath with instructions for the agent; it can also flag request waterfalls and poorly implemented effect loops before production. - Agents contributed to Solid 2.0’s validation by finding some scheduler edge-case bugs; the release went through 23 release candidates, 971 PRs, and 750 closed issues, and the team waited for a full regression-free week before calling it stable.
- JetBrains Air offers a supervised multi-agent workflow inside the IDE: run agents side by side, let them use tests, inspections, and the debugger, then send line-specific review notes back to an agent.
- Clay found more improvement in harness engineering than post-training: its trained model moved closer to Sonnet/Opus performance but did not create a step change, and the team concluded the harness had not given the model enough expressiveness. For complex company/contact search, it replaced filter composition with a SQL-like search DSL, which models can generate and which also supports composability and modularity.
- Clay started with the off-the-shelf Pi harness for its lightweightness and extensibility, kept it unforked, and added most custom behavior through plugins; its production agent uses the same CLI exposed to external agents, letting first-party agent evaluation improvements feed back into the CLI. Its agent had about 20 skills and 8–10 tools, with more than 85% of tool use through bash; it also wrote Python and bash scripts and manipulated CLI output files. To reduce latency and token use, the team identified repeatedly loading skill files as an inefficiency and considered preloading skills into the system prompt, alongside smaller optimizations.
- Clay’s evals evolved from testing individual capabilities, to launch-milestone orchestration buckets, to user-journey evals broken into subtasks—for example, inbound lead routing. The team found the middle approach awkward because coverage percentages did not translate clearly into user experience.
- For debugging long traces with hundreds of tool calls and large bash commands, asking a coding model to diagnose the raw trace worked poorly; first asking it to summarize each bash command worked better. The team identified failure-mode clustering and quantification as a useful way to prioritize fixes by frequency and impact.
Huntley’s rule of thumb is that open-weight models are typically three months behind proprietary labs.
ThePrimeagen calls Google’s Neon “the best model” and describes it as “insane.”
Tom Osman says he used Treg’s /doomscroll skill to have AI binge X, Reddit, Hacker News, YouTube, and TikTok in the run-up to Anthropic’s IPO . Jason Zhou linked to it as a tutorial for using AI /doomscroll around the IPO .
Kent C. Dodds says his agents make hundreds of mini-programs daily, compared with his own occasional creation of a couple of scripts—an example of using agents for high-volume production of small tools.
A coding agent recommended against using Effect.ts for Kody; the post asks what it got wrong but gives no rationale, and links PR #5450.
- Peter Steinberger says OpenClaw landed a fix for the Responses API prompt-caching bug in its embedded harness; he notes most harnesses had not yet updated.
-
The bug affects GPT-6.2 Astra and Sol: top-level
instructionscan remain uncached despite diagnostics reporting a cache hit. A workaround is to place the system prompt in a leading developer message; the cited report says cache hit rates fell from 97.4% to 46.4% when moving from GPT-6.1 Sol to 6.2 Sol.
Jason Zhou introduced the open-source /doomscroll skill: give it a topic and time window to find worthwhile posts across X, TikTok, YouTube, Reddit, Hacker News, and Product Hunt; he says his agent reviewed 272 GTA-6 tips and easter eggs for $0.02. Try the demo or inspect the repo.
Geoffrey Huntley described managing “OpenAI dots” through bots, using one bot per virtual employee, and combining automated supplier-risk management with a daily-updates briefing bot.
- The article links attacks on infrastructure firms to abliterated, no-safeguards models and says an attacker used Railway’s downloaded monorepo to develop a plan to reach AI-provider keys within six hours—compared with days or weeks for a human; the attacker resumed probing the same night Railway evicted them.
- Railway responded by replacing shared build machines with a throwaway VM for each build, adding fine-grained access controls for sensitive CI components, and requiring 2FA or biometrics for important credentials. The article’s broader advice: assume attackers may have source-code access, and alert and respond to suspicious activity within minutes.
Theo cautions that benchmarks show capability but provide an incomplete picture of it, a useful caveat when evaluating models for coding-agent work.
Riley Brown linked a broadcast labeled “Building with Google Neon and Claude Fable / Opus”; the post provides no workflow or implementation details.
I'm gonna miss coding
- Theo reports that Claude writes more than 90% of his code. When the agent gets something wrong, he says he either directs it to correct the issue or writes that part himself.
- Theo endorses treating prompting as a search for inputs that produce the desired code without hand-editing; he compares it to speedrunning and argues that automating more work creates room to experiment with more ideas.
- In a sponsored segment, Theo identifies GitHub rate limits and slow runners as bottlenecks, then pitches Depot: Docker-layer caching for development, CI, and agents; DepotCI, which he describes as faster and easier to parallelize than Actions; and planned Depot Code, a GitHub-API mirror intended to work with Depot CI and sandboxes for immediate code access.