Agent Workflow in a Minute · intermediate

Workflow in a Minute: Execution

Execution is the agent actually editing and running commands—bounded by the plan, permissions, and verification gate.

Last verified: 2026-09-06

In one sentence

Execution is when the coding agent implements the approved plan inside agreed permissions and stops at verification.

Why it matters

Plans do not ship. Unbounded execution also fails—by expanding scope. The execute phase must be energetic and fenced.

How it works

During execution:

  • follow the approved file list unless a blocker forces a replan;
  • run commands that serve the task, not curiosity sprawl;
  • keep commits coherent if committing mid-task;
  • stop for approval on irreversible actions;
  • enter the verification loop when implementation claims done.

If reality contradicts the plan, replan explicitly. Silent plan drift is how agents “helpfully” rewrite your system.

Example

Implement only steps 1–3 from the approved plan.
If you need files outside the list, pause and propose an updated plan.
Then run npm test.

Agentic coding use

Orchestrators often split planner and executor roles. Even in one agent, you should preserve the same phase boundary in prompts and permissions.

Watch out

“While I’m here…” edits are scope creep with good manners. Ban them in the task contract unless cleanup is an explicit second task.