Investigate and plan, with writing switched off.
Build is the working mode: it investigates, edits files, runs commands and verifies the result. Plan produces a written approach instead, and the restriction is enforced at the tool layer rather than asked for in a prompt.
Ask ML.ai Code…
Build
Edits files, runs commands, verifies the result.
Plan
Investigates and returns a plan. Cannot edit or write.
How it works
Plan is genuinely restricted, not a suggestion. The edit and write tools are denied to it, and a stricter command budget applies.
The tools are denied, not discouraged
Plan has its own system prompt, and the edit and write tools are refused outright. Writes, deletions and inline code execution are refused.
Verification still runs
Test runners and typecheckers are allowed rather than prompted, because verification is exactly what planning needs.
Auto-approve cannot widen it
With auto-approve on, a command the safety check denies in Plan stays denied. A Plan turn that deleted a directory would not be Plan.
Switch per turn
The mode selector sits in the composer bar and applies to the following turn, so you can plan, review, then switch to Build to execute.
The ask that works, and the one that doesn’t
The ask that fails in Plan mode
"Fix the registration validation bug and run the relevant tests." Plan can investigate and describe the fix, but the edit itself is refused outright, so this doesn’t work as asked.
The ask that fits Plan mode
"Explain how registration validation works, then propose a fix as a plan." This is exactly what Plan is for: investigation and a written approach, nothing touched.
You review the plan
The written approach comes back with no source file changed. Verification tools ran if needed: Plan allows test runners and typecheckers, because verification is part of planning.
You switch to Build to execute
The mode selector applies to the next turn, so you plan, review, then switch to Build to actually make the change described.
Fix the registration validation bug and run the relevant tests.
Edit refused in Plan mode
Investigate and describe the fix instead: switch to Build to apply it.
Fits Plan mode
Explain how registration validation works, then propose a fix as a plan.
Worth knowing
Plan mode restricts the tools, not the consequences. Running a test suite still writes coverage files, snapshots and caches.
Verification commands are trusted by name: npm test is treated as verification because of how the script is spelled, and nothing at this layer sees inside a project-defined script.
Questions worth asking
Related
Agents
Choose a subagent for the composer and it stays selected until you remove or replace it. Each one is built for a different kind of work, and only general carries out changes end to end.
Read moreBackground runs
A delegated job runs in its own session. Your conversation stays usable, and the run survives the panel being closed, hidden or reloaded.
Read moreTry ML.ai Code today, or talk to us about what is next.
Install the editor agent on your own machine, or book a call to talk through your team's workloads.