30 stories
OpenAI just made the engine under Codex an Apache-2.0 platform — codex exec, the SDK, and the JSON-RPC app-server. The pitch is one number: harness changes alone took the same model from 13.3% to 38.3% on ARC-AGI-3 with 6x fewer tokens.
Alibaba finally opened a Max-class Qwen — and the 2.4T checkpoint downloads at 4.89 TB, three times Kimi K3 for a smaller model, because Qwen shipped BF16 where Moonshot shipped pre-quantized 4-bit.
A nameless model with a 1M-token context dropped on OpenRouter and OpenCode, free for a week, and nobody has claimed it. The forensic trail — a leaked Java stack trace, error code 1214, 30/30 tokenizer matches — points hard at Zhipu’s next GLM.
Meta's first terminal coding agent launched with a benchmark chart it loses — and a design worth stealing: background agents that persist all session, sub-agents fanned into isolated git worktrees, and a 12–21x discount if you let Meta train on your code.
The three metrics every team reaches for first — lines written, acceptance rate, self-reported time saved — are the three that reward an agent producing more code than you needed.
Your personal setting is being ignored, or the repo's is — and the reason is that the direction of precedence is not the same for every mechanism. settings.json lets the repo beat your laptop; skills let your laptop beat the repo; CLAUDE.md lets nothing beat anything; permission rules refuse to…
The programmatic hooks you pass in options.hooks, not the shell hooks in settings.json — and the difference matters, because by default both fire in the same run. We registered a callback on all 31 events and ran it until we had the thing nobody publishes: the measured firing order.
The agent gets one shot, on a machine with no browser and nobody to ask — which changes which failures you get. We ran four headless arms on one real failing test: the locked-down run with no allowlist was denied Bash and Edit, wrote nothing, exited 0, reported subtype "success" and cost $0.383 —…
The one job everyone hands an agent first, and the one place a wrong answer looks exactly like a right one. We gave Claude Code the same 7-line function twice — once with the implementation, once with only a spec — then broke the code nine ways to see which suite noticed.
Seven questions a security review actually asks, answered from current documentation rather than a trust page — including the three where the honest answer is "nothing stops that".
The prompt everyone types is the weakest one available — it names a symptom and asks for a diff, and the cheapest way to satisfy that is to make the symptom stop appearing.
The thing people reach for the moment they want to stop approving every command — and the one fear it does not remove. A real Dockerfile, a default-deny egress firewall measured before and after, and the four things that break first: a build-arg token that shows up in docker history three times…
A migration is not one task, it is a thousand nearly identical ones — and that changes how you drive it. The pilot-then-rules-then-batches shape, and the part everyone skips: a codemod beats the agent on the mechanical 80%.
Codex is the one terminal agent whose sandbox the kernel enforces — Apple Seatbelt on macOS, bubblewrap plus seccomp on Linux (Landlock is now the legacy path).
The window is the constraint everything else follows from, and file reads are what fill it — about 31% of a representative session, against 1% for your own prompts. What /compact and /clear each do, and the part nobody warns you about: compaction does not bring everything back.
Google's terminal agent is the only one of the three with a free tier that works without a card — 1,000 model requests a day on a plain Google sign-in. The three auth paths and what each actually meters, why GEMINI.md silently ignores your AGENTS.md, the nested settings.json shape, MCP and…
The subscription-vs-API crossover is a computable number, not a matter of taste — and the tokens go almost nowhere people assume. We run a 40-turn Opus 5 session at $13.20 uncached and $2.86 warm, work the break-even for Pro, Max 5x and Max 20x, price the same session across four models at a clean…
The notify hook has exactly one event, hands you the payload as an argv argument rather than on stdin, and can never block a tool call. The verified payload shape, a runnable script, per-run vs global wiring, and when to reach for lifecycle hooks instead.
Showing 18 of 30 stories
Load more storiesGet the top AI news and tools straight to your inbox every week.
Weekly · AI news
Join 15,000+ AI enthusiasts
