How do Git worktrees and durable handoffs keep AI coding agent sessions on track?
In brief
Give each concurrent coding-agent session its own Git worktree and branch, with one live session per checkout. Keep intent, decisions and handoff notes in project files. Preserve unfinished work on its own branch so a fresh session can resume without reconstructing the chat.
- Use one worktree per concurrent editing session.
- Preserve the branch, checks and remaining work.
- Test whether a fresh session can resume.
The problem: sessions end before the work does
Coding agent session management preserves progress when context fills, a process stops or a different agent takes over. Git worktrees provide separate working directories and branches; durable handoffs preserve the task’s intent and remaining work.
Sharing a checkout risks mixed commits and overwritten edits. Isolated work without a handoff leaves the replacement session guessing what “done” means.
What you may have tried
- Reusing the same chat: retains conversation until its context or process is lost.
- A shared checkout or stash: convenient, but concurrent edits and stash ownership can become ambiguous.
- Worktrees alone: isolate working files, but do not preserve decisions or isolate shared Git storage, services and databases.
Use one live session per checkout. Preserve work on its owning branch, record checks and blockers, and confirm the previous session stopped before reusing its workspace.
How Arcadia keeps a session tied to its Action
Arcadia isolates coding-agent sessions in worktrees and carries the current action and its context in repository files. The current Action belongs to a Plan serving the Project’s Outcome. That connection lets a replacement recover the reason for the work, not just its files.
Suppose a signup repair passes its behavior check but still needs keyboard verification. A handoff identifies the branch and candidate revision, the passing check and the outstanding criterion. The next session resumes that Action instead of treating the previous agent’s summary as completion.
The benefit is continuity without rebuilding the conversation. Integration and shared services still need coordination. Ask a fresh session to locate the task and reproduce its evidence from project files before increasing parallelism.
Sources and further reading
- Anthropic: effective harnesses for long-running agents — progress records and handoffs across sessions.
Current capabilities and supporting evidence
- Agent session managementWorking
Isolated worktrees per action, across Codex, Claude Code and OpenCode.
Supporting evidence - Plans, actions and a work pointerWorking
Governed documents with one active plan and one current action.
Supporting evidence
Capability snapshot: . See what changed →