Ephemeris.
13 August 2026
Agents wrote a third of the merges. Humans still sign every one.
Vercel put the AI SDK's issue tracker on a conveyor belt: agents in isolated sandboxes classify each report, probe it with a failing test, implement, self-score for risk, and open the pull request. A person reviews the evidence chain and approves the merge — that step never moved. Four weeks in, the open-issue count is falling for the first time in months.
- Merged weekly, agent-authored
- 25–35%
- Issues closed by the factory
- 70–80%
- Open issues, late June → early Aug
- 1,022→ 844
Redis answered in 1.29 ms. The replacement takes 5.13, and that exposed the real bug.
The warm pool of standby containers behind every Vercel build kept its billing records — the only ledger of which deployment gets charged — in a cache that could evaporate. Moving to DynamoDB made each query two to three times slower, which broke a supply loop nobody knew was serialised. The fix was concurrency, and the telemetry then showed the loop had been stalling under Redis all along.
206 million behaviour checks in one day.
Cloudflare is retiring the single-moment bot score. Precursor watches a session end to end from the client side, and the numbers explain why: across 73,438 domains in twenty-four hours, suspicious behaviour usually surfaced mid-session, after the visitor had already been scored human. Sessions now flip between human and agentic mode, so the useful question is what this request is trying to do.
PrecursorClient-side continuous evaluation — 206M events across 73,438 domains in 24 hours.
BotBaseA directory of every agent seen, verified or not, so operators can sort by purpose.
Adaptive IntelligenceA self-adjusting engine replacing versioned models — no customer migration per release.
Landing later this yearRandomised block/challenge/allow to poison retry logic, AI Labyrinth, and queues for welcome automation.
Your Copilot bill now itemises cache reads.
The AI usage report downloadable from billing settings breaks every model down into input, output, cache-read and cache-write tokens beside the credits each consumed. That is enough to trace a line item back to the behaviour that caused it — and enough to prove whether your prompt caching is actually earning its keep. Available to Copilot Business and Enterprise admins, and to individual users.
| Column | What it counts | Why you care |
|---|---|---|
| input | Prompt tokens sent | Context bloat |
| output | Tokens generated | Verbosity cost |
| cache read | Served from cache | Savings, realised |
| cache write | Written to cache | Warm-up overhead |
| ai credits | Charge per model | The invoice line |
Strip the thirty AI tells, then hand it fifteen of your own posts.
A seven-step recipe for a writing skill that sounds like you rather than like a model. Start from the open-source Humanizer skill, which catalogues roughly thirty giveaway patterns, then run your own archive through an agent and have it name your sentence lengths, punctuation habits, and the words you never use. The output is one markdown file with three sections, and no code.
- To
- Anyone whose drafts read like everyone else's
- Re
- Building a voice skill from your own archive
- Encl
- One .md file, metadata header, allowed tools
- Clone Humanizer — about thirty AI-text patterns, documented.
- Baseline a draft through it and note what personality it strips.
- Collect 15–30 of your own texts; drop anything an editor touched.
- Have an agent describe your rhythm, punctuation, and absences.
- Write three sections: hard bans, encouraged, three exemplars.
- Feed back concrete corrections each cycle, update the file.
- Add a second profile per genre once the first one holds.
That's today.
Thursday, 13 August 2026. Eighteen feeds read, five stories filed. The rubric, in order: tools you could adopt this week, creative software, dev tools and agentic coding, privacy and security when the signal is high, research with a practical kernel. Nothing older than thirty days, no story that has run here before.
Also read — The Batch · The Zvi · Import AI · Jeff Bullas · OpenAI · Anthropic · Sentry · PostHog · Fly.io · Next.js · @seeallochnaya · @denissexy · @rvnikita_blog · @ProductsAndStartups