Repos for robots: when your codebase has an audience of agents

Bun shipped a porting guide written explicitly for coding agents this week, and a public leaderboard now ranks repos by how friendly they are to Claude Code, Cursor, and Codex. The hidden audience has a name.

Two signals from this week show coding agents are no longer just users of code — they're a target audience that codebases now get tuned for. Bun published a docs/PORTING.md guide explicitly written for coding agents helping with an experimental Zig-to-Rust port, and a public leaderboard now scores GitHub repos by how friendly they are to specific agents.

Simon Willison flagged the Bun guide on May 5, 2026, noting it reads like architecture documentation written for a colleague — except the colleague has no intuition about your codebase. Bun later confirmed the port is an experimental branch, not a decision to switch languages. Either way, the PORTING.md exists, and its primary reader is a model.

The Agent Friendly Code leaderboard formalizes what platform teams have been quietly optimizing for. It scores repos on signals like AGENTS.md / CLAUDE.md presence, CI configuration, test coverage, and dev-environment reproducibility — broken out per agent across Claude Code, Cursor, Devin, GPT-5 Codex, Gemini CLI, Aider, OpenHands, and Pi (as of May 2026). "Agent-readable" is becoming a measurable repo-quality dimension, comparable to bus factor or test coverage.

Willison also observed on the Heavybit podcast this week that vibe coding and agentic engineering — once distinct workflows — have begun to blur in his own work. That convergence makes the broader picture coherent: if you're shipping software that agents will use to ship more software, the codebase needs to be legible to them first.

For builders, the practical move is to stop treating AGENTS.md as a clever hack and start treating it as a first-class artifact. Mistral's Leanstral cloud agents, launched in March 2026, push further by trying to remove human-in-the-loop code review entirely — which only amplifies the need for unambiguous repo metadata. And Ollama's new Claude Code integration, which recommends at least 64k context tokens, moves agent cost off the cloud bill while exposing how much these tools lean on context infrastructure to do anything useful.

The bottom line: If your repo doesn't have an AGENTS.md by the end of Q2 2026, you're optimizing for a workflow that's already legacy.

Sources