# Orchestration Source: https://docs.ideharmony.com/orchestration Orchestration is what turns a set of agent panes into something that can take a goal and come back with reviewed work. It is also the least settled part of the product. > [!WARNING] > Modes in this section vary in maturity. Some are implemented and exercised; > others are designed ahead of being proven end to end. Check > [feature status](/docs/reference/feature-status) before depending on any of > them. ## The separation that explains everything | Component | Owns | Consequence | |---|---|---| | **Client** (Desktop or TUI) | Presentation and your decisions | Closing it does not stop a run | | **Daemon** | Orchestration state: plans, tasks, delegation, artifacts | Single durable writer | | **Session host** | Provider processes and PTYs | Owns what is actually executing | The TUI is a client, not the orchestrator. This is why a run continues when you close a window, and why two clients can watch the same run. ## An execution route "Which model ran this?" is not one question. A complete route is: | Element | Meaning | |---|---| | **Harness** | The agent program being driven | | **Provider** | The vendor behind it | | **Account** | Which of your identities with that provider | | **Model** | The specific model requested | | **Effort** | Reasoning depth, where the provider exposes it | | **Tool set** | What the agent is permitted to call | | **Platform** | Where the process runs | | **Billing class** | Whose money is spent, and under which plan | Any of these can differ between two steps of the same plan. A route shown before consent is a route you can refuse. > [!IMPORTANT] > If projected cost or billing route cannot be determined, a run should block > rather than proceed. Consent to spend you cannot see is not consent. ## Modes | Mode | What you approve | Maturity | |---|---|---| | **Plan only** | Nothing runs; you get a plan | Most conservative | | **Guided** | Each significant step | Middle ground | | **Automatic** | One scoped consent for a reviewed plan | Requires trust in the plan | Start with plan-only against a real goal. Read what it proposes before you allow anything to execute. ## The lifecycle
Orchestration lifecycleA goal becomes a plan, which is checked against policy, then routed to a provider and model per step. Work is delegated to agents in isolated worktrees, verified against evidence, integrated by an explicit decision, and only then complete. Two stages carry gates: routing blocks when projected cost or billing route cannot be determined, and verification requires evidence rather than an agent's claim of success. Consent is given once, after the plan and its routes are shown.YOU DECIDE HEREGoalPlanPolicyRouteDelegateWorkVerifyShipGATE — COSTGATE — EVIDENCEblocks if cost or billing route is unknownagent claims are not evidenceConsent once, after the plan and its routes are visible — not per action while tired.
Two stages are gates rather than steps. Routing is where cost becomes real; verification is where "done" is either earned or merely asserted.
Goal → plan → policy → route → delegate → work → verify → integrate → complete. Two places deserve your attention: - **Route**, because that is where cost and provider choice become real. - **Verify**, because agent-reported success is not evidence. ## Isolation Delegated work runs in its own git worktree so parallel agents do not collide in the same files. See [worktrees](/docs/concepts/worktrees). ## What not to assume > [!NOTE] > Queued is not delivered. Delivered is not accepted. Accepted is not completed. > A busy-looking fleet can be entirely blocked. Read task state rather than > inferring from activity. ## Related - [First orchestration run](/docs/getting-started/first-orchestration-run) - [Worktrees](/docs/concepts/worktrees) - [Agent-to-agent messaging](/docs/a2a) - [Company mode](/docs/orchestration/company-mode)