Instructions for AI agents contributing to this repository.
A curated list of resources, patterns, and templates for building reliable AI agent harnesses. The primary artifact is README.md.
- All content in
README.mdis in English. - Each resource entry follows the format:
- [Title](URL) — 1–2 sentence note explaining why it's worth including. - Notes should be opinionated: explain why the resource matters, not just what it is.
- Do not add resources without a note.
- Sections in
README.mdare organized by the problem being solved, not by vendor or model.
Resources that address a specific harness engineering problem:
- Context delivery and compaction
- Tool design and MCP
- Planning artifacts and task decomposition
- Permissions and sandboxing
- Memory and state management
- Verification and evals
- Agent loop structure and orchestration
- General AI/ML papers not specific to agent harnesses
- Model capability benchmarks unrelated to harness design
- Product marketing without technical substance
- Tutorials on using a model (vs. building scaffolding around it)
Files in templates/ are reusable starting points. When editing templates, preserve the comment structure — the comments are the value.
Before any PR:
- All URLs are reachable
- All entries have a 1–2 sentence explanatory note
- No section has more than ~10 entries without a clear reason
-
README.mdrenders correctly as Markdown