Ephemeris
PRODUCTION
Today's issue is about the week AI tooling grew up: durable agents, instant apps, custom silicon, and five million people using it for real work.
Click the link. The next page is already there.
Next.js 16.3 (preview) borrows the single-page-app trick to kill the dead beat after a click. Await data on the server and you now choose how a route navigates: Stream a loading shell, Cache it with 'use cache', or Block when you genuinely want the wait. An instant() Playwright helper and a Navigation Inspector keep instant routes from regressing — slow ones become errors in dev.
Prefetch, rethought: not one request per link — one shell per route, cached and reused across every link that points there.
Agents that survive the restart.
AI SDK 7 ships a WorkflowAgent (@ai-sdk/workflow) for durable, resumable execution that lives through process restarts, redeploys, and human approvals that arrive an hour later. A standardized reasoning option maps to each provider's native setting, and the SDK adds the plumbing a long-running agent needs to be safe in production.
Step four failed. Walk the other three back.
Cloudflare Workflows now take a rollback on step.do() — a compensating action paired with each forward step (debit, then credit). When a run terminates mid-flight, rollbacks fire in reverse start-order, even across parallel steps, and survive an engine restart by replaying durable step history without re-running side effects. The one rule: make handlers idempotent.
OpenAI's first chip does one job — inference.
Jalapeño is OpenAI's first Intelligence Processor, built with Broadcom from design to tape-out in nine months — what they call the fastest high-performance ASIC cycle yet. It's a purpose-built inference part, not a repurposed trainer: a large compute chiplet plus HBM, tuned for the data-movement and memory bottlenecks of reasoning and agentic workloads. Lab samples already run GPT-5.3-Codex-Spark at target power.
- Design → tape-out
- 9 months
- Built for
- LLM inference
- Perf / watt
- Beats SOTA*
- First deploy
- End of 2026
Every department hired the agent.
OpenAI's usage data marks the crossing from demo to daily work. Research requests are up 56× since November 2025, support 32×, engineering 27×. The work got longer, too: by May, 80.6% of requests mapped to tasks a person would spend over 30 minutes on, 70.2% to over an hour. Knowledge workers — Legal, Recruiting included — are now ~20% of Codex users and growing three times faster than developers.
That's today.
A thin docket with thick signal: the week durable workflows, instant apps, and purpose-built silicon all landed at once, and the usage numbers caught up. Five picks, one rubric — what a senior engineer or founder could put to work this week.