We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Huntley: stop risky agent actions by removing the capability
Geoffrey Huntley's talk "the answer to ai is engineering" (in a blog post and on YouTube) is the most reusable material this period. His central rule: "The best way to stop a model from doing something is to remove its ability to do it." Agents push toward their goal and do destructive things along the way. So he asked Claude to "use Nix and create a custom overlay that removes force push." The patched git is now used on his laptops, in agent sandboxes and by anything that depends on git .
The rest of the approach, from his nix-demo repo:
- One environment file for everything. A single devenv/Nix file builds developer laptops, CI/CD, agent sandboxes and production Docker images. This avoids running a separate agent-sandbox setup that drifts from the others . CI compiles patched software once and publishes it to a cache, and the other environments download it from there .
- Handling security incidents. He asks Claude whether his fleet is affected. It then applies patches and updates refs, dependencies and pins . For temporary patches, he prompts it to "update the AGENTS.md, and in seven days' time, get rid of my patches." The next agent to use the machine removed them once upstream was fixed .
- Root access only on throwaway machines. His agent loops run with sudo on disposable Nix dev machines, and NixOS tests run work against infrastructure that matches production . He never lets agents run as root on production hosts or connect to production databases. For reproductions, he copies the database and tightly limits outbound network access .
- Don't let the loop accept bad code. Pre-commit hooks block pushes that don't compile. Static or semantic checks enforce architecture rules, such as SQL code never referencing the API . Property-based tests find edge cases that unit tests miss . He works at Antithesis, which makes Bombadil (property testing for web apps, which returns a failing example as a reproduction) and Hegel (a Rust core with layers for each language) .
Gemini 4 Carbon in real use: slow, hallucinates, but good for security fixes
Theo's reports add real-use data to last period's CursorBench result:
- Speed. One task took 22m21s with Carbon on low effort, against 9m33s for Opus 5.5 on high . On identical code-review tasks it was 4x slower . He blames mostly Google's servers: 11 seconds before the first response, 20x longer than Fable . He also says its tool calls don't batch or run in parallel, and that it needs a "code mode" like Opus, Astra and Fable use to run actions in one call .
- Hallucinations. False claims across his evals fell from 36% with the previous Gemini model, Argon (which he calls "Argan"), to 18%. One Rust architecture review still had 23 false claims out of 79, where most models have zero or one . He says not to use it for code review or codebase analysis. He saw it invent database column names, line numbers and architecture details .
- Cost per task. Theo disputes claims that Carbon is cheaper. He says Opus 5.5 at xhigh effort costs less than Carbon at any effort level on real tasks .
- Where it did well. He gave it a vulnerability report for T3 Code. It proposed separating known-machine flows from raw-URL flows instead of a heavy cryptographic redesign. He then had it get a "council review" from other models and fix their feedback, and reviewers scored the result 7–9 out of 10 . He warns that his early access may have made the model's safety refusals less strict for him .
- If you use Gemini through another tool: use the Google-approved ACP bindings. Some of his hallucinations came from an old binding that didn't pass line numbers to the model .
What Carbon produced: a Windows-install fix for T3 Code that took 14 hours and ran to 4k+ lines where about 500 should do. He had Fable clean it up , and Fable's PR was "significantly smaller and better" . Nicholas Moy, more positive, calls Carbon "very good." It is publicly available in Antigravity with generous limits for Ultra subscribers .
Fable 5.2: new prompt wording and an arena result
-
Prompt update. Anthropic has changed the recommended "This codebase is yours" sentence to:
You are the owner and long-term maintainer of this codebase, and its design is up to you. - Agent Arena ranks Fable 5.2 (Max) #1 across 4.7k agentic sessions, at +30.4% net improvement against +19.5% for Opus 5.5 . Median cost per task fell from $7.7 to $7.3, while Opus costs $5.5 and Carbon $1.5 per task. One row of the post's own table labels the top entry "Fable 5.1" .
- Matthew Berman's team: Fable built a complete multi-level Fall Guys clone. Opus "builds you one level and goes, 'Hey, is this good enough?'"
- Simon Willison used leftover Fable credits to have it analyse 244 tools with tree-sitter. It tagged the JavaScript APIs, libraries and web API hosts each tool uses, and those tags now let you browse tools.simonwillison.net by tag .
- Riley Brown gave one Claude Code prompt to make a 4K Blender video. Fable wrote Python to build the scene, rendered test clips and graded them itself, and added timed sound . It rented 10 RunPod GPUs and rendered 1,329 frames in 52 minutes for about $6 .
Smaller items
-
Find plaintext keys: grep your agent configs for
sk_,ghp_,sntrys_andxoxb-. Each hit is a key your agent can read (guide) . - Limit self-review loops: LangChain's GTM agent re-reviewed pages that were already fine, sometimes dozens of times. A rule to "ship once one review finds no real defect" cut a one-pager from 18 minutes to 7 .
- Antigravity browser use is rolling out. You choose which accounts it logs into, downloads wait for your approval, and the agent can run several browser tasks at once .
- llm 0.37.2 fixes two ways your OpenAI key could be sent to a third-party OpenAI-compatible provider .
- aws-emulator is AWS's official emulator for testing locally against AWS services. A swarm of agents wrote it in Rust, with 29,684 commits in its first three weeks. It needs Docker on macOS and Windows .
- Monorepos: Matt Carey argues against them for agents unless you'll rebuild git tooling. He suggests a CLI or MCP that lets agents search your organisation's code without a checkout .
- Local models: Omarchy 4.6 will share a model on one machine with others on your tailnet . DHH runs GLM-5.3 Flash on two DGX Sparks . Theo says open-weight models are under 0.5% of T3 Code traffic .
- In an experiment to receive aircraft ADS-B at 1090 MHz, the SX1276 could not tune to the frequency, while the SX1262 could but lacked OOK support. He asked an AI to trace RSSI; when it could not sample signal strength quickly enough to detect the carrier, he redirected it to investigate direct sampling or chip reprogramming.
- He used an RTL-SDR as a live reference, testing the SX1262 when an aircraft signal was actually present, and suggested synchronizing sampling on the first of the two symbols. His approach paired his hardware and signal-processing context with the AI’s strength in algorithms and DSP; he said the experiment’s success was still uncertain. The SX1262’s internal ARM CPU can accept opcodes in place of its firmware, potentially enabling functions supported by the hardware but not initially provided by the factory firmware.
- For repeatable AI work, he recommends providing a few Markdown files with project conventions and adapting the setup to the specific problem, rather than treating MCP, skills, or loops as essential; understanding the problem and the AI’s capabilities is more important.
-
Huntley uses one
devm.nixdefinition for developer laptops, CI/CD, agent sandboxes, and production Docker images; Nix overlays let him patch tools such as Git to remove force-push capability and carry custom versions through transitive dependencies. -
For incident response, he asked Claude whether his fleet was affected; the agent updated refs, dependencies, and pins and applied needed patches. He also asked it to add a seven-day cleanup reminder to
agents.md, so a later agent removed temporary patches after upstream fixes landed. - For agent loops, he gives agents sudo/root only on deliberately designed throwaway developer machines—not production hosts or databases—and uses NixOS tests to spin up multi-machine network setups and test the stack against infrastructure defined as code. He adds deterministic backpressure with pre-commit compile and static/semantic checks, and recommends property-based tests to explore edge cases and produce failing counterexamples.
-
Use
devenv/Nix as a single source of truth for developer laptops, CI/CD, agent sandboxes, and production Docker images to reduce environment drift; Nix caching and remote builders let CI build and validate patched software once for those environments to reuse . To constrain agent actions, Huntley asked Claude to create a Nix overlay that patches Git to disable force-push, applying the patched Git in laptops and agent sandboxes; the principle is to remove a risky capability from the underlying system rather than just prompt the agent not to use it . -
For security incidents, ask Claude whether the fleet is affected, then have it apply needed patches and update refs, dependencies, and pins; for temporary fixes, add an
AGENTS.mdreminder to remove patches after seven days so the next developer or agent handles cleanup once upstream is fixed . - Run agents as root only in a throwaway Nix development machine, where they can inspect system state, and test their work against production-like infrastructure defined as code; never run them as root on production hosts or connect them to production databases—duplicate the database for reproduction and restrict outbound network access .
- Keep agent loops from accepting bad output with deterministic backpressure: pre-commit checks can block pushes when code does not compile, and static or semantic checks can enforce architecture boundaries. Add property-based tests to explore generated inputs and verify invariants, such as reversing a string twice returning the original or a customer refund occurring only once . For browser-app properties, Huntley describes Bombadil exercising JavaScript/TypeScript apps under randomized conditions and returning a counterexample when a property fails; he also names Hegel, a Rust core with language-specific layers for property-based testing .
- In Theo’s evaluations, Gemini 4 Carbon’s false-claim rate fell to 18% from Argan’s 36%, but it still made 23 false claims in a 79-claim Rust architecture review. He cautions against using it for codebase analysis or review, where it invented database columns, line numbers, and architecture details; he also found it slow and would not choose it as his default coding model.
- Its strongest use for him was security incident response: he supplied a vulnerability report, asked for a fix, and iterated on the proposal. For a T3 Code peer-link issue, Carbon recommended separating known-machine flows from raw-URL flows and tightening URL retention, rather than a heavy cryptographic redesign; Theo then had it build a solution, get a quick review from other models, and address their feedback. Reviewers gave it scores ranging from 7 to 9 out of 10. He noted that early-access status might have affected how strongly the model’s safeguards applied.
- A reusable workflow he tested in T3 Code was having Carbon use the tool to get feedback from other models and then continue the task; on a Windows remote-update investigation, reviewers rated its implementation 8–9/10, though one reviewer was Argan. For Gemini use through another tool, he says to use a Google-permitted ACP integration: an older T3 Code binding failed to pass line-number information, which he says contributed to hallucinations, and DeepMind audited and improved the bindings.
- For a Blender marketing video, the presenter gave Claude Code a single brief; the agent wrote Python to build the scene, created a website and placed its screen recording into the Blender scene, animated the camera, lighting, and character, tested renders and graded them, then generated music and sound effects. It rendered 1,329 frames in 4K using 10 RunPod GPU pods (two 5090s and eight 4090s) in 52 minutes for nearly $6, compared with over five hours on his MacBook Pro.
- For agent-assisted video editing, he supplies footage from a camera or folder and gives a concrete brief—for example, vertical 4K, a fast flying motion, and sound effects—with optional reference material or style guidance. He recommends version control: keep prior cuts, prompts, and comments, then ask the agent to use comments on the video to make the next version; he describes v2.video as a free tool for this workflow.
- In his app-building trials, vague prompts were the main cause of mistakes; he recommends clearly describing the app and letting the agent test it on the computer and iterate until satisfied.
- Riley reports using Claude Code to turn a single prompt into a 4K Blender marketing video: the agent planned the piece, wrote Python to build the scene, created a website and captured its screen for an in-scene display, animated the camera, lighting, and character, rendered and self-graded test sections, then generated and timed sound. It rendered 1,329 frames using 10 RunPod GPU pods (two 5090s and eight 4090s); Riley says this took 52 minutes and nearly $6, versus more than five hours on his MacBook Pro.
- For app building, Riley found that lazy prompts caused mistakes; he recommends clearly describing the desired app and letting the agent test it on a computer and iterate until satisfied.
- For agentic video editing, he emphasizes version control. He presents v2.video as a free tool that tracks versions, prompts, and comments on specific parts of a video; those comments can guide Claude’s next revision.
- A useful coding-agent workflow is to plan with a stronger model, then delegate tasks to smaller models to stretch quota; the video describes Cognition’s fast coding model, post-trained from Inkling Small, as intended for coding subagents.
- For open-weight coding-agent experiments, Recursive Intelligence’s Ripple 65B Preview is reported at 80.5% on Terminal Bench 2.1, with up to a 1M-token context and pricing of $0.20 per million input tokens and $0.60 per million output tokens. The speaker says it can run locally, including on a MacBook or DGX Spark.
Anthropic says Mythos got Claude running on AMD in three days with “basically no human intervention,” porting or writing almost 100% of the kernels; the company contrasts this with previous new-chip-platform integrations that took dozens of engineers months.
Simon points to Anthropic’s Fable 5.2 prompting guide: framing the codebase as the model’s own (“This codebase is yours”) is intended to make it less cautious about improving existing code—merging duplicated code and making related fixes instead of only extending existing structure or reporting nearby problems. In Anthropic’s small coding-task tests, this wording led to merging repeated code in about half of attempts, versus under one tenth without it; results varied by task, and the effect held inside a long system prompt. Simon later noted that Anthropic updated the recommended wording to “You are the owner and long-term maintainer of this codebase, and its design is up to you.”
- Matthew Berman’s team said Fable 5.2 built a multiplayer Fall Guys clone in one try, hosted on here.now; it worked in the stream demo, where the room filled at eight players.
- In their comparison, Fable handled long tasks to produce a complete game with multiple levels, while Opus produced one level and asked whether it was enough. They also reported Fable got a separate build right on the first try where Astra took five or six iterations, and surfaced ideas they had not anticipated.
Google introduced Gemini Decision 1, a decision-model API supporting yes/no, choice, and rubric-scoring questions. It accepts text, images, audio, video, and PDFs; Jev is limited to text/JSON, while OpenAI supports text and images. Pricing is $0.10 per million input tokens with free output, compared with Jev at $0.042 per million. In a video test, Gemini returned probability scores for yes/no answers (0.923 for a robot, 0.050 for a dragon), rather than the 1.0/0.0 values Jev and OpenAI return.
LLM 0.37.2 fixes a security issue in its OpenAI-compatible endpoint support: a configured OpenAI API key could be sent to another provider if the key named by api_key_name had never been stored, or if a responses: true model was run with -o chat_completions 1.
AWS’s official aws-emulator lets developers test locally against clones of AWS services without provisioning a paid API key. The Rust project was written by what Simon Willison calls a “huge swarm of agents” and had 29,684 commits in its first three weeks, though the post gives no details about the agents’ workflow. It currently requires Docker on macOS or Windows, so it is not yet a single-binary option for broader test-suite integration.
To make 244 vibe-coded tools easier to browse, Simon Willison used remaining Fable 5.2 credits to analyze each tool’s code with tree-sitter, identifying JavaScript APIs, libraries, and web API hosts; the resulting categories power tag-cloud browsing on the homepage .
-
Use one Nix
devenvfile for developer laptops, CI/CD, agent sandboxes, and production Docker images to reduce environment drift; Nix overlays can apply package patches across those environments, with CI rebuilding transitive dependencies and publishing them to a shared cache. -
For security incidents, Geoffrey asks Claude whether the fleet is affected, then has it update pins and apply patches; he adds a dated cleanup instruction to
AGENTS.mdso a later agent removes the patch after upstream fixes the issue. He describes this overlay-plus-expiring-note pattern as his standard response. -
Give loop agents
sudoonly on disposable Nix machines; never give them root on production hosts or connect them to production databases. Use copied data and restrict outbound network access, and test changes on production-identical infrastructure. - Gate agent loops with deterministic checks: pre-commit compilation, static and semantic architecture rules, and property-based tests for edge cases missed by unit tests. He names Bombadil for property-based web-app stress testing and Hegel for property testing across languages.
ThePrimeagen says he has been “vibing” hard and now every change takes three hours, which he finds frustrating—a negative productivity signal from this workflow.
AI Engineer Code Summit advertised a free livestream with Lydia Hallie of Anthropic on Claude Code best practices and N. Serrino of Gimlet Labs on “10x faster AI,” scheduled for 9 AM PT.
Jason shared a builder’s workflow for choosing a product idea using demand evidence: provide skills, budget, and available time (or leave them blank); the tool generates about 24 ideas, checks them against live Treg search data, then returns a winner, two runners-up, the rationale, build difficulty, and first steps. The described workflow has AI write only idea names and one-liners; metrics such as search volume, ad prices, keyword difficulty, top results, backlinks, and Reddit data come from Treg, and missing values are left marked missing rather than filled in. To use it with a coding agent, create a repo, set up the coding platform, and give the agent set up treg - https://treg.to/llms.txt, then build and ship; the article describes Treg as providing these data tools through one key, with per-call pricing shown up front and no added markup.
Google introduced Gemini Decision 1 as a frontier decision model in the Gemini API . It natively supports image, audio, video, and PDF, with a 256K context window and “1K questions”; pricing is $0.10 per 1M input tokens with no additional cost, and Google says it can be tried free in the API .
ThePrimeagen says monorepos have always been an anti-pattern . In the linked agent-era argument, Matt Carey says teams should avoid monorepos unless they are willing to reimplement Git/GitHub, and instead give agents a lightweight CLI or MCP to search organizational code and docs without checking out the repository .