Foundations.
Today's issue is about the layer underneath: open specs for orchestrating agents, Git as a primitive for them, a base model with a new attention shape, and the renderers and runtimes everyone else builds on.
DeepSeek V4 brings a new attention.
The Pro variant is 1.6T total parameters with 49B active and a long-context attention scheme called CSA/HCA; the Flash sibling is small enough to be cheap and useful. If you build on open weights, this is the next baseline to test against.
| Variant | Total params | Active | Input · Output ($/M) |
|---|---|---|---|
| V4-Pro | 1.6T | 49B | $1.74 · $3.48 |
| V4-Flash | 284B | 13B | $0.14 · $0.28 |
An open spec for conducting Codex.
Symphony turns issue trackers into always-on agent systems: an open-source orchestration spec that says how a queue of tasks fans out across Codex workers, who owns what, and how results land back. Wire it once, leave it running.
# a Symphony task in plain YAML trigger: issue.label.added=codex/auto workers: 4 handoff: - plan → implement - implement → review - review → open-pr on_fail: reassign & retry(2) # the spec; not a SaaS, not a product
Versioned storage that speaks Git.
Artifacts gives an agent a real git remote: clone, branch, push, fork from anywhere. Spawn ten million repos for ten million sessions, hand off a URL to any standard git client, throw the rest away. The hard part of long-running agents has always been state — this is one of the cleanest answers yet.
The stack underneath very large models.
Cloudflare's foundations post is the rare inference write-up that sits below the model: custom kernels, scheduler, KV-cache layout, and the assumption that the network is the bottleneck. The bar chart isn't the point — the chart of who owns each layer is.
WebStreams, ten times faster.
Vercel rewrote WebStreams from the spec down and shaved 10–14× off Next.js rendering benchmarks. Less object allocation, less server overhead, fewer ceremonies between chunks. If your SSR is hot, this is the kind of upstream win you get for free on the next runtime bump.
$ bench next-render --before --after ───────────────────────────────────────── before ▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮▮ 4180 ms after ▮▮▮ 312 ms ───────────────────────────────────────── 13.4× faster · server overhead -78%
Building faster by building less.
A long, generous tour of how Turbopack tracks the dependency graph at function-call granularity, memoizes everything, and rebuilds only what changed. The same idea Salsa, Adapton, and Bazel circle around — applied to the front-end build, which has wanted it for a decade.
The fastest computation is the one you don't perform. Incremental computation is, more than anything, a discipline of remembering enough to know what's safe to skip.
What GitHub learned letting agents drive.
A team inside Copilot Applied Science spent a quarter doing real engineering with agents in the lead. The takeaways aren't about prompts — they're about how teams change shape when the bottleneck moves from "writing the code" to "deciding what to ship next."
- 01Specs become the artifact. The PR is downstream.
- 02Reviews shift from style to intent — the agent already wrote it cleanly.
- 03Tests are negotiation. They're how you tell the agent it's done.
- 04Stand-ups become roadmap meetings. The day-to-day is in the queue.
Ticket to PR in four minutes, one dollar.
A working demo, not a deck: an autonomous pipeline that reads an issue, opens a Docker sandbox, runs the Claude Agent SDK end-to-end, and pushes a pull request — total wall-clock around four minutes per ticket, total cost about a dollar. The interesting part is how thin the orchestration layer can be.
That's today.
Eight stories about the layer underneath today's product surface — model attention, agent storage, build graphs, the runtime. The picks tilted toward primitives because the primitives had a better week than the products did.