Planning

The Agentic Orchestrator handles each request at the lowest execution tier that fully covers it. Most requests never need a plan, since each plan costs the user a review before work starts.

How It Works

• Choosing a tier. Anything answerable from the conversation gets a direct answer. Three or fewer steps run as ad hoc tool calls. Longer work gets a plan, especially when results feed later calls or the user should approve the approach first.
• Plan review. The user sees a short task list and replies to approve, request changes, or cancel. Changes come back as a revised plan. An ambiguous “no” doesn’t cancel; the Orchestrator asks.
• Sign-off. Destructive, irreversible, or broad-reach actions need explicit sign-off first: plan review for planned work, a direct confirmation for ad hoc calls, unless the user just asked for exactly that action.
• Concurrency. Independent ad hoc calls in one response run concurrently as a batch. In a plan, phases and batches run in order, and the steps within a batch run concurrently.
• Waiting. The active plan’s status and goal stay in context every turn. When the session needs the user, it suspends instead of holding a request open and resumes on the next message. Asynchronous results from external systems fold back in when they arrive.

Configure

• Instructions can’t turn off plan review or the sign-off rule.
• To show the user exact values before a specific action runs, set the action activity’s Confirmation Policy in the plugin’s Conversation Process.

Related

• Memory: context, session state, and retention.
• Tool Orchestration: how tools are called.