Solutions
Hand off tasks between AI coding agents
Forkbench's task board is a durable per-Thread file, not a chat transcript, so a task one agent writes today, complete with its plan, notes, and any discussion that resolved a question, can be claimed and finished tomorrow by a different agent from a different vendor entirely. The board is free on every plan, which is deliberate: it is how agents sharing a job keep out of each other's way, not the thing Forkbench is for. What it is for is the scope around it, where the keys and notes one job needs are the ones an agent working that job can reach, and no others.
Most agent work dies with the session that produced it. Close the terminal pane, or quit the app, and the plan an agent was building lives only in scrollback you'll never revisit. And even the tools that do keep some memory tie it to one agent or one vendor: what Claude Code wrote down, Codex can't read, and a new CLI session rarely knows what the last one figured out.
The board is what makes this possible: a per-Thread JSON file with atomic writes and a synchronous flush at app terminate, so it survives a closed pane, a quit app, even a reboot. Tasks carry no vendor field, so nothing on the board says "Claude only" or "Codex only": any agent in any pane can claim any open task. Closing a pane hard-expels that agent's claim immediately, so a stalled task never sits locked waiting for a lease to expire; the 120-second lease is only a backstop for an agent that's still alive but stuck. Progress notes carry where the work actually is, the worktree and the branch and the approach taken, so the next agent doesn't have to go looking for it. And when two agents disagree about how something should be done, that argument becomes a discussion on the task rather than a message that scrolls away. A task can't be marked finished while a discussion on it is still open, so whatever got settled is written down before the work is allowed to close. None of that includes your keys: the task carries its plan and its notes, but the credentials an agent may use are granted to the Thread, not to the task, so handing work to a different agent never hands it a key you did not already grant that workspace.
How it works
- 1An agent researches or plans, then writes what it found, and a proposed next task, straight to the Thread's board.
- 2Agent-authored tasks land pending, so you accept it once, from the Mac, the web, or your phone.
- 3Close the pane, or quit Forkbench entirely: the board is a per-Thread file with atomic writes, flushed to disk at app terminate, so nothing on it is lost.
- 4Come back whenever, run a different agent, even a different vendor's CLI, and have it claim the accepted task, read the discussion that resolved it, and continue exactly where the first agent left off.
Straight about the guarantee: The board doesn't verify the next agent is equipped to finish the job, only that the notes and discussion are there if it is. That judgment call is still yours.