AutoMode / How it works
How AutoMode picks a model and effort
AutoMode chooses the cheapest model and effort that are still adequate for each prompt. Two Haiku judges read the prompt, code sets a quality floor for risky work, and a cost check decides whether a switch is worth it. The judge can never lower a floor. The decision takes about a second.
How a route is decided
- Judges give evidence. Two Haiku judges assess the prompt. A failed judge never blocks your prompt.
- Code sets a floor. Rules check for risk, such as security work or a migration. No model sets this floor.
- The optimizer picks a candidate. It takes the cheapest eligible pair of model and effort.
- Authority resolves the route. A user override or workflow role is final. The floor then lifts the candidate. The switch check holds or switches on cache cost, and a hold never drops below the floor.
The ladder
| Rung | Model | Effort | Example prompt |
|---|---|---|---|
| T0 | Haiku | Low | What does the -r flag do in cp? |
| T1 | Sonnet | Medium | Rename this helper and update its tests |
| T2 | Sonnet | High | Refactor the importer to stream rows |
| T3 | Opus | High | Move sessions from cookies to signed JWTs |
| T4 | Opus | Max | Any prompt where a focused attempt already failed |
These are examples, not live output.
What counts as risky work?
Code sets a risk band from S0 to S3, using the task text, the file paths, the module count and the unknowns. The band only goes up. A prompt reaches S3 if it matches security or credentials, migrations or destructive operations, concurrency, external contracts, or payments and sensitive data. S3 work always gets an Opus review.
What happens to each prompt
- Chat and non-code requests get a direct answer.
- Code changes with a clear goal: Sonnet does them, and must show a passing test or build before it says done.
- High-risk changes: after the checks pass, one Opus reviewer checks the diff.
- Open-ended work runs through the workflow below.
The workflow for big jobs
| Stage | What runs |
|---|---|
| Risk band | Code, not a model. The band only goes up |
| Discover | Haiku. One agent for S1, three for S2 and S3 |
| Plan | Opus. Workstreams, acceptance criteria and risks |
| Build | Sonnet. Workstreams that share files never run together |
| Verify | Sonnet runs the commands. Code decides pass or fail from exit codes |
| Review | Opus reads the diff. Mandatory for S2, S3, failed checks and unfinished work |
How do I override a route?
Start a prompt with am: to steer that one prompt. For example, use am:opus, am:sonnet, am:haiku, or am:low to am:max. Your override is final. Use /automode off to stop routing for a session, and /automode on to start it again.
Where does it run?
Routing runs in the terminal, the Code tab of the Desktop app, the VS Code extension, claude -p and the Agent SDK. The bar above the prompt appears in the terminal and the Desktop app. In VS Code, use /automode for the same facts. AutoMode does not run in a WSL session of the Desktop app.
How well does it work?
The judge picked an accepted effort on 74 of 75 labeled cases. This measures the label, not cost. AutoMode has no published cost result yet, so we quote no saving. See how other routers compare.
Install AutoMode
Paste these into Claude Code, one at a time, then run /reload-plugins.
/plugin marketplace add mralexivy/AutoMode
/plugin install automode@automodeIt is free. It needs no key and no server, and it starts on its own.