Build and run / Harnesses

Harness safety and cost

Updated 28 Oct 20264 min read

A harness runs AI agents with real access, often without a person watching each step. These controls are enforced by the platform itself, not by asking the bot in its instructions.

Test without changing anything

ControlWhat it does
Dry runRuns one bot on its own. Every outside change (a pull request, a message, a cloud change) and every task or memory write is saved as a draft instead of being done. Works on a bot that is still turned off.
Shadow modeRuns the whole harness. Bots really run, and outside changes are saved as drafts you can review on each task.

Start every harness in shadow and test one small piece of work before you switch it on.

Limit what bots can do

ControlWhat it does
Permission ceilingEach harness type has tools its bots can never call. An AppSec harness can't delete user accounts or cloud resources; a GRC harness can't merge or push code.
Tool approvalThe harness policy can mark a tool as needing a person, such as merging a pull request, or as never allowed. A run that tries it is stopped until someone approves it on the task. One approval allows one use.
Fail closedIf the platform can't check a tool call against the harness's rules, for example during a short outage, anything that would change something is refused. Reads still run.
A person takes overAssign a task to yourself and its bots stop working it. Hand it back and they continue.

Approval choices follow risk: a step that can be undone, such as opening a pull request, usually needs only a notice; a step that can't, such as merging, deploying or revoking access, needs approval. See Autonomy and approvals.

Limit how much runs

ControlWhat it does
Visit limitsEvery loop, such as fix then verify then fix again, has a maximum number of visits. When it is reached, the task goes to a person.
Concurrency limitsA stage can be limited to a number of bots running at once.
Run time limitEach bot has a limit for one step of work, and a deadline for its whole run.
GroupingA scan's findings are worked together, so one scan starts one fix run, not one per finding.
Monthly budgetA harness can have a monthly spending limit. When it is reached, new bot runs wait instead of starting.

Schedules belong to the harness

The harness sets the schedules of its scanners and other scheduled bots. When you switch a harness on, its schedule is written onto those bots, and Optimus tells you beforehand if that changes a bot's current schedule. While the harness is on, a bot's schedule can only be changed in the harness, not on the bot: the bot page refuses the change and says which harness owns it.

A bot that works a stage, such as the fix bot, never has a schedule of its own. On its own clock it would run with no task. The platform refuses a harness that sets one up that way.

Stop it

  • Disable on the Harnesses page pauses the harness and turns off all its bots, including scheduled scans.
  • Pause holds new work; resume continues where it stopped.
  • Each run can be cancelled from its execution, by a person or by Optimus. A cancel is recorded on the task, with who cancelled it.

Know what happened

Every action a harness run takes outside the platform, and every cancel, is recorded under What it did on the task, with links to what was created. The platform writes these records from the tool calls themselves; a bot can't write, change or skip them. Secrets in the tool's inputs are hidden.

Together with each task's journey and timeline, this gives an auditor the full story: what came in, which rule moved it, which bot ran, what it changed and who approved it.

Untrusted input

Findings, webhook payloads, pull request text and ticket text come from outside and are passed to bots as untrusted data. For harnesses that read attacker-controlled content, such as reported phishing, use separate bots: a triage bot that reads with read-only tools, and a response bot that holds the privileged tools behind approval.

Next steps

Need a hand with this?Our engineers can walk through it with you on your own stack.