Claim before edit
Stop AI coding agents overwriting each other's files
Running two or three agents on one repo? Have each one claim files before editing, so the second agent learns about the first before it writes anything.
The problem
Agent A rewrites a file while agent B is halfway through changing it. Neither knew the other was there, and one set of edits silently wins.
What people try
- Tell each agent which folders it owns in the prompt
- Nothing checks it. The boundary drifts the moment a task needs one shared file.
- Separate worktrees per agent
- Files stop clashing on disk, but the collision moves to merge time, after both agents have built on it.
- A lock file or TODO list agents are asked to update
- Agents skip it, and a stale entry looks the same as a live one.
With Concord
01
Claim
Before editing, the agent calls start_work with the files, modules and domains it expects to touch.
02
Detect
Concord compares that claim with every active task and reports same-file, same-directory and shared-module overlaps.
03
Split
The second agent picks other work or messages the owner to divide the file. Neither has written a line yet.
Result
Agents find out about each other at claim time, while backing off costs one prompt instead of a rewritten file.
Common questions
- Is this a file lock?
- Not by default. A claim is a declared intent that other agents see, and overlaps are reported rather than enforced. In Claude Code you can opt into an edit gate that blocks writes to files another active task has claimed.
- What if an agent doesn't know which files it will touch?
- Claim modules or domains instead, and add files as they become clear. Re-claiming merges the new scope into the task and recomputes overlaps.
- Does it work if the agents are in different tools?
- Yes. Claude Code, Codex, Cursor, Gemini CLI and other MCP clients all claim through the same local workspace.
Further reading
Give your agents shared work-state.
npm install -g @concord-ai/concord-mcp