Ephemeris · Issue 080 Thursday · 09 Jul 2026 · Zürich

Ephemeris.

“One engineer can do a lot more today than a year ago.”

— Bun's team, on rewriting 535,496 lines of a JavaScript runtime from Zig to Rust in eleven days, with an AI at the keyboard. Today's issue: the machines took the engineering shift.

Ephemeris · Issue 080 01 / 6 · Bun · via @seeallochnaya
Dev tools · Runtimes

Rewritten by machine, in eleven days.

Bun's team ported its JavaScript runtime from Zig to Rust to end a class of use-after-free crashes — driving Claude Fable 5 through dynamic, adversarial multi-agent review. The full suite passes on every platform, and use-after-free is now a compile error, not a 2 a.m. page.

535,496lines · 1,448 files
11days · May 3–14
$165Kin API tokens
Repeated Bun.build() memory: 6.7 GB → 609 MB · ~20% smaller binary · 2–5% faster HTTP
Ephemeris · Issue 080 02 / 6 · Vercel
Dev tools · Agents

An agent you can let near production.

Vercel Agent investigates incidents, fixes failing builds, and reviews pull requests — but only inside a plan you approve. It runs as its own identity, read-only by default, executing generated code in ephemeral Firecracker microVMs against real project copies.

“Even if something goes wrong, Vercel Agent can only act inside the plan you approved.”

● REC   00:03:00 · bad deploy detected → rolled back

Ephemeris · Issue 080 03 / 6 · OpenAI
Research · Evaluations

The eval hit a noise ceiling.

OpenAI re-audited SWE-Bench Pro — the coding benchmark it had urged the field to adopt — and found a large share of tasks broken in ways that make scores unreliable. It is now formally retracting that recommendation.

SWE-Bench Pro · public split · 730 tasks
Signal sourceBroken tasksShare
Automated flagging pipeline200 / 73027.4%
Human review · 5 engineers249 / 73034.1%
Effective noise ceiling—~70%
Ephemeris · Issue 080 04 / 6 · Cloudflare
Infra · Distributed systems

Consensus without a leader.

Cloudflare's Meerkat swaps leader-election for QuePaxa: any replica can drive agreement, and concurrent proposals cooperate instead of colliding. Across 330+ data centers where latencies swing unpredictably, that removes the single timing-out point Raft depends on.

10×throughput vs. Raft & Multi-Paxos under a hostile network

“Leaders in our proof-of-concept clusters constantly fail, and the cluster keeps operating with no increase in error-rate.”

Ephemeris · Issue 080 05 / 6 · GitHub
Dev tools · Automation

Docs that write themselves, then wait for you.

GitHub's Agentic Workflows are GitHub Actions with an AI model as the work-item processor — authored as markdown with an English prompt and compiled to constrained Actions. Aspire's pr-docs-check turns every merged feature into a drafted, correctly-branched docs PR. Nothing auto-merges.

on: merge → main / release/* ▸ read the product diff ▸ decide if docs are needed ▸ draft in the house style ▸ resolve target branch (milestone → release) ▸ open DRAFT PR → route to the feature's reviewer

“Where are the docs? The honest reply is usually some variant of ‘behind.’”

Ephemeris · Issue 080 06 / 6 · GitHub · via @TochkiNadAI
Dev tools · Self-hosted

An agent you run by messaging it.

right-agent turns a Telegram chat into an agent that uses your own Claude subscription — no third-party API keys. Every agent runs in its own sandbox; every credential lives outside it, so a compromised run can misuse a tool but never walk off with a token. Rust, self-hosted, one command to start.

$ curl -LsSf https://raw.githubusercontent.com/onsails/right-agent/master/install.sh | sh $ right up ▸ agent online · telegram bound ▸ sandbox: isolated · credentials: external ▸ v0.4.2 · try /mcp · /providers █
End of issue 080 Back to top ↑

That's today.

Issue 080 · 9 July 2026. Six picks for the engineer or founder with ten minutes: an AI-driven runtime rewrite, an agent scoped to touch production, an honest audit of a coding benchmark, leaderless consensus, self-drafting docs, and an agent you can self-host.

Fed today by Bun · Vercel · OpenAI · GitHub · Cloudflare, with a Rust agent surfaced via @TochkiNadAI.
Also scanned: The Batch · Zvi · Import AI · Jeff Bullas · Anthropic · Sentry · PostHog · Next.js · Fly.io · Telegram (seeallochnaya, denissexy, rvnikita_blog, ProductsAndStartups).
Rubric: adoptable AI tools · creative & dev software · agentic coding · security when the signal is high · research with a practical kernel.