Separate planning and reasoning from agent action

Unlock the complete study guide.

Bundled into the existing Developing in Agentic AI Systems premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

Planning and action are separate stages with a gate between them
Configuring planning and implementation roles through their agent profiles
Custom handoffs: passing the plan to an implementation agent
Reviewing the VS Code Plan agent's plan before Start Implementation
Where the Plan agent saves its plan
Plan and implementation model settings are not a gate
Copilot CLI: plan mode, plan.md and the approval step
GitHub Copilot upgrade: a reviewable Markdown record at each stage
Exam reasoning: find the gate, then check what bypasses it

The four gate surfaces: where each plan lives, what holds its gate, and what opens it

Surface	Where the plan lives	What keeps the gate closed	What opens the gate
VS Code custom agent	The chat response, shaped by the agent body	Read-only research tools without the agent tool, a handoff with send: false	Edit tools, or send: true on the handoff
VS Code Plan agent	/memories/session/plan.md, cleared when the chat ends	Iterating on the plan before Start Implementation	Selecting Start Implementation
GitHub Copilot CLI	plan.md in the session folder	The plan workflow waits for approval; plan mode blocks project-file edits as a safety net, not a guarantee (some shell commands and calls to Model Context Protocol tools pass)	Approving the plan, or starting with --plan --mode autopilot
GitHub Copilot upgrade	assessment.md, upgrade-options.md, plan.md and tasks.md under .github/upgrades/{scenarioId}	Guided mode pauses at each stage boundary	Automatic mode, or saying "continue"

Decision tree

Cheat sheet

Give a planning-only agent read-only tools rather than implementation capabilities.
Define the structured planning deliverable in the custom agent's Markdown body.
Prevent automatic subagent invocation without hiding the agent from human selection.
An empty agents list stops a custom agent from starting subagents.
Use /plan or plan mode to request planning before implementation.
Copilot CLI stores its structured implementation checklist in the session's plan.md.
Inspect and edit the CLI plan before authorizing implementation.
The CLI plan workflow waits for approval before implementing.
Plan-then-autopilot is unsuitable when a human must approve the planning-to-action transition.
Validate the VS Code Plan agent's implementation and verification steps against the requirements.
The VS Code Plan agent's automatic plan save is session memory, not a durable repository document.
Use a non-sending handoff when the developer must inspect the implementation prompt.
Hand off a finalized plan through Start Implementation, carrying the conversation context.
Configure planning and implementation model defaults independently in VS Code.
Use assessment.md to review the upgrade's diagnosed compatibility problems and scope.
Review upgrade-options.md for the confirmed decisions that planning should implement.
The upgrade plan converts assessment findings and confirmed options into an implementation specification.
Validate project interdependencies before removing work from an upgrade plan.
The execution-stage task list pairs concrete changes with success checks.
Use guided upgrade mode when each stage requires review before the next stage starts.
Set upgrade targets and execution policy before the staged workflow begins.
Choose upgrade order according to dependency depth and feedback needs.
Unresolved upgrade problems require intervention rather than an assumed successful stage.

Unlock with Premium — includes all practice exams and the complete study guide.