Instruction files
AGENTS.md vs CLAUDE.md when you run several agents
Codex reads AGENTS.md, Claude Code reads CLAUDE.md, Cursor reads its rules. Keep standing instructions in one place, and keep live work-state out of all of them.
The problem
Three agents, three instruction files, and they drift. Then the files fill up with task notes that were only true last week.
What people try
- Copy-pasting the same rules into each file
- The copies diverge the first time someone edits only one.
- Putting current task status in AGENTS.md
- Every agent reads it as truth, long after it stopped being true.
With Concord
01
One source of rules
Keep shared rules in AGENTS.md. Claude Code's CLAUDE.md can import it with a line like @AGENTS.md, so both read the same text.
02
Rules, not status
Instruction files hold conventions that change rarely. Who owns what, and what was decided today, belongs in task state.
03
Let setup add the rest
concord setup adds a short block to each client's instructions telling the agent when to claim, record and hand off.
Result
Instruction files stay short and stable. Live state stays current because it has its own home.
Common questions
- Does Claude Code read AGENTS.md directly?
- Claude Code reads CLAUDE.md. Importing AGENTS.md from CLAUDE.md is a simple way to share one set of instructions.
Further reading
Give your agents shared work-state.
npm install -g @concord-ai/concord-mcp