How do I manage multiple AI coding agents without losing track?
In brief
Manage multiple AI coding agents with isolated workspaces, one owned task per session, explicit priorities, approval boundaries and completion checks. Keep those records outside the agents’ chats so a new session can continue from the project rather than from your memory.
- Give each session one bounded task.
- Keep priorities and dependencies outside the chat.
- Check the integrated result before adding more agents.
The problem: parallel work multiplies coordination
Managing multiple AI coding agents means keeping ownership, dependencies and proof clear across simultaneous sessions. Extra capacity helps only when the changes serve the same intended result and can be integrated safely.
Otherwise, you become the missing coordination layer: reconstructing priorities from tabs, finding branches and deciding whether “finished” changes actually work together.
What you may have tried
- Assigning work in separate chats: starts quickly, but priorities and dependencies remain in your memory.
- A shared backlog: exposes tasks without necessarily identifying the next permissible Action or unresolved decision.
- Worktrees and separate builds: protect working files and check individual changes; they do not eliminate integration conflicts or isolate shared services.
Give each session one bounded task and workspace. Record dependencies, approval boundaries and an integrated completion check before adding another agent.
How Arcadia connects several sessions to one Outcome
Arcadia maintains plans and actions with one work pointer, records decisions and approval gates, and checks completion evidence against criteria written before work starts. A Project supplies the Mission and Outcome; a Plan groups Actions toward a Milestone. Sessions execute that bounded work while Decisions preserve human authority and evidence establishes completion.
For a signup change, the implementation Action and documentation Action can belong to one Plan. The documentation depends on verified interaction behavior, so the dependency explains why it remains a draft. Isolated worktrees keep edits separate; a combined check still establishes readiness.
The benefit is visible coordination that survives a session change. Start with two tasks: can another session find their priorities, branches and remaining approvals without asking you to reconstruct them? Increase parallelism when that handoff works.
Sources and further reading
- Git worktree documentation — isolated working trees and the repository state they share.
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 - Decisions and approval gatesWorking
Consequential calls are recorded and wait for a human answer.
Supporting evidence
Capability snapshot: . See what changed →