Daily Eval

Thu 5 Nov 2026

A weekday email for software engineers working alongside AI agents. Five minutes: what changed, whether it matters for you, and one thing to try.

Sample.

Bloomberg's LLM migrations opened 2,333 PRs and landed 729

  1. 001Bloomberg landed 729 migrations from 2,333 LLM-assisted pull requests
  2. 002Kimi K3.1 Pro now served by more zero-data-retention providers on Vercel
  3. 003OpenAI generated patches for 948 Linux kernel privilege-escalation bugs
  4. 004Ripple-65B Preview claims 80.52 on TerminalBench 2.1 as an open-weight model
  5. 005OpenAI is buying Pierre to expand Codex code review tools

How.

  1. 001Every weekday, a pipeline reads 57 sources: 53 feeds, Hacker News, Hugging Face daily papers, Lobsters and new GitHub repos.
  2. 002Today, 155 items became 14 shortlisted, then 5 bullets you can skim.
  3. 003Each story gets TRY, WATCH or SKIP. The one thing leads; Safe to ignore names the hype.

Who.

Written by John Young, a CTO who runs coding agents in production.

He builds flywheel, an open-source agent orchestrator.

jyoung.dev  /  github.com/johnayoung/flywheel

Join.

Agentic AI, evaluated every weekday. Five minutes.

2026-11-05.md

Thursday 5 November 2026

Bloomberg's LLM migrations opened 2,333 PRs and landed 729

Five bullets

The one thing

Bloomberg found writing migrations easy and landing them hard

Bloomberg reported on Automated Uplift, a multi-month effort to run LLM-assisted migrations across thousands of repositories. One campaign covered about 1,900 Python packaging repos. Another covered about 6,700 C++ build system repos. Together they opened 2,333 pull requests and landed 729 successful migrations.

The pipeline handles as much as possible with deterministic code. The LLM only makes bounded file transformations inside context computed ahead of time. Runtime telemetry, such as execution logs and usage traces, also feeds the pipeline.

Generating the migrations turned out to be the easy part. The bottleneck was the cost for owning teams to review, test and deploy them. Costs fell when the pipeline could verify changes before opening a PR. C++ compilation and test runs made that possible. Costs also fell when repos were uniform enough for one fix to carry over to the next.

This is a pattern you can copy for your own backlog. Let scripts do the finding and the context gathering. Keep the model on small, bounded edits. The skill that grows is building verification that runs before anyone has to review.

Script what you can, give the model one bounded edit, and verify before the PR opens.

Try this · 15 minutes

See how much of a migration needs the model at all

  1. Pick one small migration you keep postponing, like a packaging or build-file change.
  2. Write a script that finds the affected files and gathers the context each edit needs.
  3. Give your agent one file at a time, with only that pre-computed context.
  4. Compile and run tests before opening a PR. Note what a reviewer must still check.

Stories

Safe to ignore

57 sources read. 155 items. 14 shortlisted. 5 bullets.