Swarm

The harness behind Codex is now one API call.

The Agents SDK from April put the agent loop in your code: you ran the harness, you owned the infrastructure, you kept the sessions alive. The Agents API, in public beta since Wednesday, hands that whole orchestration layer to OpenAI — the same Codex harness, hosted, with context compaction, subagent coordination and a managed sandbox. Sessions are built to survive for days. There is no separate API fee; you pay for the models and tools a session actually burns.

Eighty-eight hours, ten thousand agents, and a 166-page proof.

On 8 September OpenAI claimed the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize problems. The run started on 1 September and finished 88 hours later, and it shipped with a Lean formalisation a machine can check. Read it for the method, not the trophy: this is what a research swarm costs and how its output is made auditable.

Write HTML. Render video. Same frames every time.

HyperFrames is HeyGen's Apache-2.0 renderer that turns HTML, CSS and animation into MP4 through headless Chrome and ffmpeg. GSAP, Lottie, Three.js and plain CSS animations all work, and the output is deterministic — same input, same frames — which is what makes it safe to put in CI. It ships 20 agent skills and a component registry, so an agent can author motion graphics the way it authors a web page. Denis Sergeev cut a reference-matched motion piece with it in seventeen minutes this week.

Node compatibility is on by default, and the ceiling is 64 mebibytes.

Cloudflare rewrote the Workers module registry, and two of the changes remove reasons you were not deploying there.

8,913 articles across 70 fake newsrooms, in 20 languages.

Anthropic's September threat report is the most concrete public accounting of what people are doing with a frontier model when nobody sanctions it — named clusters, real counts, and the detection signals that caught them.

215.8 ms to 19.1 ms, by putting an index inside the shard.

Vercel's CDN used to fetch routing metadata one path at a time. On a deployment with hundreds of thousands of paths, every deploy reset the cache and every request paid a miss. The rewrite groups metadata into ~200 KB JSONL shards with sorted paths and Base64 offsets, so one fetch warms many paths and a binary search over fixed-width pointers parses only the record you asked for.

Your bundler assumes two-thirds of visits are a single page.

Turbopack merges chunks to cut requests, and the merge maths is weighted by a guess: 0.67 of sessions end on the first page. Sam Poder's write-up shows the guess, then hands you the dials. Next.js 16.3 adds firstPageLoadPriority (start from your bounce rate), priorityRoutes, and clusters for routes visited together — plus generateComponentChunks, which emits un-merged chunks alongside merged ones so the runtime picks whichever is cheaper given what the browser already holds. turbopackSharedRuntime saves a blocking request and ~10 KB per navigation; turbopackCjsTreeShaking finally shakes CommonJS.

One signature is 2,420 bytes. UDP budgets 1,232.

1.1.1.1 now validates DNSSEC signed with ML-DSA-44, the NIST post-quantum scheme, so the internet can stress-test the awkward part at resolver scale. The awkward part is size: a single signature is 38× an ECDSA P-256 one and blows the DNS UDP budget on its own, forcing truncation and a TCP retry. Fragmenting UDP is not an option worth taking.

The migration hazard is subtler than the size. Zones publish classical and post-quantum signatures side by side, and RFC 6840 lets a validator accept any single valid path — so an attacker forges the breakable algorithm and the post-quantum signature never gets asked. Plan for larger responses and a TCP or DoH path now.

One agent found the autograder bug. Twenty-seven minutes later, the swarm had it.

DeepMind pointed 100 Gemini 3.1 Pro agents at 71 problems from Formal Conjectures, gave them a bulletin board, direct messages and a shared library, and told them not to cheat. Thirty-seven problems in, prover-theta found a hole in the autograder. The library that made them collaborative made the exploit contagious, and the remaining 34 problems fell to it. Jack Clark's Import AI 472 covers the paper; the operational reading is that a shared memory is an attack surface, and that the population split matters more than the average.