~70%
Typical payload reduction over priming
Case study · Multi-agent
A market where the buyers weren't human.
Anthropic ran an experiment where Claude agents negotiated 186 transactions for staff personal items, $4,000 in total. The notable finding wasn't that it worked — it was the spread. Agents on Opus made about $3 more per trade than agents on Haiku. Bigger budget, bigger margin: a fairness footnote for any AI marketplace that lets users bring their own model.
Transactions186
Gross volume$4,000
CounterpartiesAgent ↔ Agent
Opus vs Haiku+ ~$3 / trade
ConclusionCompute = leverage
Privacy · AI infra
No retention, by contract.
Vercel's AI Gateway now enforces zero data retention end-to-end: prompts and completions can't be logged, cached, or used for training by any provider on the path. For teams shipping AI features into regulated workloads, this turns a procurement question into a checkbox — pick the gateway, get the guarantees, skip the bilateral DPAs.
Path · ZDR
appgateway
gatewayprovider
[no log · no cache · no training · enforced]
Web platform · Compression
Compression catches up to the bots.
As agents do more of the browsing, the web's classic "load it once, cache it forever" assumption breaks: every page is now hit by a fresh client with a cold cache. Cloudflare's shared-dictionaries support lets sites prime compression across requests, so the second visitor — bot or human — gets the diff against a primed baseline instead of the whole document. Small change, big ratio.
1
Shared dictionary, many cold-cache fetches
RFC 9842
Implemented spec, not invention
Manifesto · Dev tools
You should write an agent.
Thomas Ptacek's argument is unfussy: agents are the cloud's next foundational primitive, the way containers were a decade ago. If you want to understand where infrastructure is going, the fastest way is still to write something that runs against it. He gives the case for stopping consuming agents and starting building one — even if it's small, even if it's bad.
1 · Pick a loopPick a tight, useful loop you do by hand — code review, log triage, calendar wrangling — and write the dumbest version.
2 · Ship the harnessDon't model-shop. Pick one model and obsess over the harness around it. The harness is where the work is.
3 · Run it in prodAn agent you don't run in prod is a script with extra steps. Real load is what teaches you what's broken.
Operations · Security debt
A defense outlived its attack.
GitHub publishes the kind of post you rarely see from a platform of its size: a study in security mitigations that stayed in place long after the threats they targeted were gone. The lesson is uncomfortable for any ops org with a long memory — every guardrail eventually pays a tax, and the only honest fix is to instrument them and review them on a clock.
FromPlatform Security
ToAll teams running long-lived mitigations
ReDefense lifecycle, end-to-end
ActionAdd expiry / observability to every block rule.
Framework · Agent tooling
Build for the agent's point of view.
The Next.js team built an in-browser AI agent, ran it for a quarter, and then turned it off. The retro is the post: what they learned was that proper agent support isn't a chat box bolted onto the docs — it's MCP endpoints, machine-readable errors, and tooling that assumes the operator can't see the screen. They've changed how the framework ships in light of it.
| Surface | State | Why it matters | |
|---|---|---|---|
| 1 | AGENTS.md in create-next-app | Shipped | Agent priors travel with the project. |
| 2 | MCP integration | Shipped | Tooling, not chat. |
| 3 | Browser log forwarding | Shipped | Agents see what the dev sees. |
| 4 | In-browser agent | Sunset | Wrong layer. Lesson kept. |
Growth · Architecture
Your lifecycle emails are wrong, structurally.
Sara Miteva's argument is structural, not tactical: behavioral email tools were built for broadcast, so any "send when user does X" rule lives one synced property removed from the truth. The fix is to move the trigger logic to where the product data lives — your warehouse, your analytics layer — and keep the email tool as a dumb pipe. Cleaner data, fewer ghost sends, easier audit.
Reel · 008 / 008Free Models · daily test
Tools · OpenRouter
Which free models are actually working, today.
Denis ships a small, opinionated thing: every day a script pings the free tier on OpenRouter, runs each model through an agent-style task, and ranks them by whether they actually answered. The output is a public leaderboard you can pin as a tab. Not a benchmark suite — a status page for the free shelf, refreshed while you sleep.
Models surveyed~30
Refresh cadencedaily
Task classagent-style multi-step
Outputranked, public
Cost$0
Fin · Issue 009↑ Back to file
That's today.
Eight stories about the unglamorous wiring being laid for agents — markets, gateways, dictionaries, harnesses, lifecycle audits. Sources that fed today's picks:
· Anthropic (anthropic.com)
· Vercel (vercel.com/blog)
· Cloudflare (blog.cloudflare.com)
· Fly.io (fly.io/blog)
· GitHub Engineering (github.blog)
· Next.js (nextjs.org/blog)
· PostHog (posthog.com/blog)
· @denissexy (Telegram)
· @ProductsAndStartups (Telegram)