Core concepts
Agent mode and Plan mode
Two ways to run the same request, and how to pick between them.
The same request can be run two ways. Agent mode works autonomously with tools until it is satisfied; Plan mode asks you questions first and then walks a fixed pipeline. Neither is better in general, and the toggle sits on the composer so you can switch per run.
Agent mode
In Agent mode the work is handed over as a goal rather than as a fixed sequence. It can read and write files, build and simulate the design, check that it synthesises, and look things up; it decides what to do next based on what the last attempt produced, and keeps going until the result holds up or it reaches its ceiling.
- Strengths. Anything where the right next step depends on what the last one produced: debugging, iterating on a failing testbench, multi-file work.
- Cost. Variable, because the amount of work varies. Every run has a ceiling on what it may spend, so a difficult one stops rather than spiralling; see When a run stops early.
- Limits. Verilog and SystemVerilog only. Project mode requires it.
Plan mode
Plan mode reads your description, asks you a handful of clarifying questions, and only then generates. From there it follows the same fixed sequence every time, which is what makes it the predictable option.
- Strengths. A well-specified block where you would rather answer four questions up front than review a wrong module afterwards.
- Language. The only mode that produces VHDL.
- Predictability. The same request tends to take the same shape every time, because the stages are fixed.
The questions step is covered in Clarifying questions.
Choosing between them
| If you want to... | Use |
|---|---|
| Write a standalone module in Verilog | Either; Agent needs less input from you |
| Write a module in VHDL | Plan |
| Edit an existing project | Agent; it is the only option |
| Chase a bug through simulation | Agent |
| Be asked about the ambiguities before code is written | Plan |
| Keep the cost of a run tightly predictable | Plan |
Where the toggle is
Two places, both writing to the same setting: the Agent / Plan pair on the composer, and the Mode section at the top of the Run setup panel. To change the mode every new chat starts in, set it in Settings under Run defaults.
Mode is not the same as intent
This toggle decides how a run works. What the run is being asked to do is a separate choice: build, debug or optimize. Debug and Optimize runs always use Agent mode, so on those the toggle is replaced by a fixed label rather than offering a combination that would be refused. See Build, Debug and Optimize.