Steering & Fine-Tuning

Autopilot is meant to fit your operation, not the other way round. This page covers every way to shape what it does — from a single on/off switch to giving each agent written guidance and connecting your own tools.

Everything here lives in the Harness panel and in Settings.

1 · Switch agents on or off

Every agent is off by default. You turn on exactly the ones you want, one at a time — there’s no all-or-nothing switch. A common path is to start with the Inquiry agent, get comfortable, then add the Case agent, and so on.

The Dream agent has no on/off switch — its housekeeping is always on and runs on a quiet schedule. It never changes your inquiries, cases, sources, or detectors, so it’s safe to leave running.

2 · Choose observe-only or managed

For each thing Autopilot can touch, you choose how much rope it gets:

You can set this at three levels, from broad to specific:

The more specific setting wins. So you can run the whole instance in managed mode but pin your most sensitive source to observe-only — or the reverse.

A good rollout: start instance-wide in observe-only, read Autopilot’s proposals for a few cycles, then switch the areas you trust over to managed.

3 · Give each agent written guidance

Toggles decide whether an agent runs; guidance shapes how it behaves. Each agent has a plain-text field where you describe your priorities in your own words — no special syntax, just write what you’d tell a new analyst.

Guidance is the everyday way to fine-tune Autopilot. It applies to that agent on every run.

4 · Set standing rules with operator directives

For permanent policy — rules that should always apply and never be forgotten — use operator directives (managed under the Memory tab). Unlike ordinary learned notes, the Dream agent never prunes or rewrites these, so they stick.

Good directives are short and absolute:

• “Never open cases for the staging environment.”
• “Treat all IBAN findings as high severity.”
• “Our primary focus is customer-data exposure.”

See Memory & System Brief for how directives differ from ordinary memory.

5 · Run it manually and steer a single run

You don’t have to wait for a scan. From the Harness panel you can start a run on demand and point it where you want:

A manual run reviews all of your open data (not just the latest scan’s findings), which makes it the right tool for a deliberate sweep or a focused question. An instruction is a temporary steer; for permanent rules use an operator directive instead.

6 · Connect your own tools (advanced)

Out of the box, the agents already have everything they need to read findings and manage inquiries, cases, sources, and detectors. If you want them to reach your systems too, you can connect external tools using the Model Context Protocol (MCP).

• Register an MCP server under the relevant Settings page, and its tools become available to Autopilot.
• You can scope which agents may use which tools, so — for example — only the Case agent can call an internal enrichment service.
• Every external tool call is gated and logged exactly like a built-in one, including observe-only enforcement.

See the MCP Server settings for setup details.

Don’t forget the AI provider

The agents that reason with a language model — most importantly the Detector agent — need an AI provider configured. If you’ve enabled an agent and it isn’t doing anything, the Setup section of the System Brief will usually flag a missing provider as the reason.

Configure one under AI Providers.

Quick reference

Next: see exactly what Autopilot did and why in Flight Recorder & Audit.