Agents & pipelines
An agent is a named instruction with its own model, thinking budget, and permission mode. Select a few from the picker and your task runs through them in order — a pipeline, where each stage hands off to the next.
The ten built-in agents#
NeuralCode ships with ten agents covering the work you do most. Each one comes pre-tuned with a sensible model, thinking budget, and permission mode — for example, /plan runs in plan-mode so it maps the work without touching files, and /review is read-only.
| Agent | What it does | Permission |
|---|---|---|
/plan | Maps the work before any code changes. | plan-mode |
/goal | Clarifies the objective and success criteria. | ask |
/review | Finds bugs, risks, and code smells. | read-only |
/test | Writes and runs tests for the change. | auto-edit |
/optimize | Improves speed and memory. | auto-edit |
/refactor | Cleans up without changing behavior. | auto-edit |
/docs | Writes or updates documentation. | auto-edit |
/security | Audits the change for vulnerabilities. | read-only |
/typefix | Resolves type errors. | auto-edit |
/pr | Opens a pull request for the work. | full-access |
The composer picker#
Open the 🧠 Agents button in the composer to browse and select agents. Selected agents show as numbered chips — the numbers are the run order. You can also just type the name (/plan, /review, …) and it's selected for you, the same way commands & skills work.
Pipeline execution#
When you send the task, it runs through the selected agents in order. Each stage executes with its own settings — so a pipeline might plan in Opus with extended thinking under plan-mode, review read-only, then write tests with auto-edit on Sonnet. The output of one stage is the starting context for the next.
- 1Pick the agentsClick 🧠 Agents (or type their names) to build the chip row — e.g.
/plan→/review→/test. - 2Describe the taskWrite your prompt as usual. It's fed into the first agent.
- 3Watch it flowEach stage runs in order with a live status; finished stages show a check, the current one spins, the rest are queued.
/pr last. A read-only /review after your edits catches problems before they reach you.Settings → Agents#
Manage agents in Settings → Agents. Enable or disable any agent, edit its prompt, model, thinking budget, and permission mode — or add your own and delete the custom ones you no longer need.
Agents
Enable, edit, or add your own. Each agent is a named instruction with its own model, thinking budget, and permission mode.