Click any tag below to further narrow down your results
Links
The author describes how juggling 20 parallel Claude Code or Codex threads wore out their focus, so they built a system where tickets live as markdown in git, an agent works the board in cycles, and the author only handles a prioritised inbox of items needing human judgement. Each ticket is closed out when it is decided, so context does not pile up in memory.
- The core problems are sequencing and context: running many agent threads at once overloads working memory, and the author treats that as the real failure mode.
- The agent researches, builds safe tickets, runs tests and browser checks, and works on branches for sensitive code, but never deploys to production. The human keeps that step and every judgement call.
- Each ticket is closed by writing the verdict back into git and returning the item to the queue, so only one decision is open at a time. The author says it is still tiring but no longer leaves them distracted or lost.
OpenAI is buying Pierre Computer Company, the small team behind the Trees and Diffs libraries already used inside the Codex app and the Code Storage service for storing code under Git. The goal is to build Git workflows suited to coding agents, with the deal still subject to regulatory approval.
- Pierre's Trees and Diffs libraries are already used in Codex for browsing files and reviewing changes, and Diffs includes an Edit mode for fixing lines directly in a review.
- Pierre also built Highlights, a fast syntax highlighter that updates as code is typed or streamed from an agent, and designed the API for Code Storage.
- Open-source Pierre projects stay open-source, Code Storage remains available to existing customers while its features move into Codex, and the two companies stay separate until the deal closes.
OpenAI built Pando, a new storage and serving backend for Git that keeps the Git client and protocol but handles many concurrent pushes and very high read volume. It now runs their monorepo and cut median push latency nearly 5x.
- Pando coordinates each branch independently and uses causal consistency instead of repository-wide ordering, so pushes to unrelated branches don't contend on one coordination point.
- It reuses packfiles across fetches rather than building one per client, with local caches hitting over 99%, which lets clones and fetches saturate 10 Gbps links on standard cloud VMs.
- In production it delivers nearly a petabyte of Git data per day, and the team contributed upstream changes to Git such as packfile-uri improvements and "explicit haves."
Anthropic built Scion, a custom version control system, to replace Git because Claude's code generation capability pushed their merge rate from 2,500 to 30,000+ code changes per day—a scale Git's distributed design couldn't handle. The system treats clients as always-online agents and moves computation to the server, ditching Git's model of complete local repositories.
- Anthropic's monorepo became unmanageable under Git: the history alone is 100 GB, and they were doing 25 million checkouts daily for CI, which would generate 600 PB of traffic if done naively.
- Scion's core design flips Git's assumptions: clients stay thin and online, the server handles all history queries and merges, and there's no single fully-materialized copy of the repo anywhere—everything is distributed and horizontally scalable.
- The system uses "changesets" (write-ahead log entries) instead of Git commits, storing only what changed relative to the parent plus content-addressed pointers to blob storage, allowing high-throughput acceptance with no packfiles or cross-server locking needed.
Agora is a system that lets AI agents collaborate on research by storing all work as immutable Git commits in a directed acyclic graph, with a scoring system that rewards verified results. Researchers tested it with 13 language-model agents working to initialize a neural network without training, and they achieved 62% of the performance gap to a trained baseline in 12 days.
- Thirteen agents working independently with no central planner published 1,703 contributions over 12 days, with the best result closing 62% of the gap between random initialization (3.39 bpb) and a trained GPT-2 baseline (1.0 bpb).
- The winning approach used bigram statistics extracted from six donor models' predictions, factorized by SVD, plus sparse deterministic edits to sublayers—no gradient updates or training data involved.
- Early gains came fast: the first eight improvements accounted for roughly 70% of total progress, but agents converged into a monoculture around one recipe until shown a visualization of their own concentration.