AGENTS.md


ID	`agent_instructions`
Category	Configuration
Features	None
Dependencies	None

Reads AGENTS.md from the session workspace and injects its content into the system prompt. The file is re-read on every conversation turn, so changes are picked up automatically without restarting the session.

Tools

None — this capability only contributes to the system prompt.

How It Works

Agent sends a message
Before processing, the system reads /workspace/AGENTS.md from the session filesystem
Content is wrapped in <agent-instructions source="AGENTS.md"> XML tags
Injected at the beginning of the system prompt (before other capability prompts)

Use Cases

Project conventions — coding style, naming rules, architectural constraints
Build instructions — how to compile, test, and deploy the project
Context injection — repository structure, key files, domain-specific terminology
Dynamic guidance — update instructions mid-session by modifying the file

Example

Upload an AGENTS.md to the session workspace:

# Project: acme-api

## Stack
- Python 3.12, FastAPI, SQLAlchemy
- PostgreSQL 16, Redis 7

## Conventions
- Use snake_case for all Python identifiers
- All endpoints return JSON with `{ data, error }` envelope
- Tests go in `tests/` mirroring `src/` structure

## Build
- `pip install -e ".[dev]"` to install
- `pytest` to run tests
- `ruff check .` to lint

The agent will follow these instructions for every turn in the session.

Notes

File path: /workspace/AGENTS.md (plain Markdown, max 32 KiB)
Re-read every turn — edits take effect immediately
If the file doesn’t exist, no instructions are added (no error)
Works with File System tools to update instructions dynamically