System Manual
Trust & Governance
How much an agent may do on its own: the trust level an admin sets, the autonomy an agent earns from its track record, the approval gates each stage relaxes, and the rules that hold at every stage.
Two controls work together
Workforce keeps two separate records of trust for each agent type in each organization. One is set by a person. The other is earned by the agent.
| Trust level (0 to 4) | Set by an admin in Settings. Every agent starts at 0, where it may only observe. |
| Earned autonomy (five stages) | Tracked automatically from the agent's decisions. It decides which approval gates the harness may relax. |
The trust level an admin sets
Each agent type has a trust level from 0 to 4. A new organization starts every agent at level 0, where the agent may only observe. A run started from the dashboard is refused while its agent is still at level 0, with a message to raise the level in Settings. The Reader and the Center are the exception, because they only read and route.
A level can only be raised one step at a time, and it can be lowered by any amount. The dashboard names the levels like this:
| Level 0 | Observe Only. The agent analyses but takes no action. |
| Level 1 | Draft Only. The agent creates drafts for a person to review. |
| Level 2 | Execute (Low Risk). The agent acts on pre-approved skill types. |
| Level 3 | Execute + Notify. The agent acts and sends a notification. |
| Level 4 | Full Autonomy. The agent acts without a notification. |
The autonomy an agent earns
Each agent also carries a trust profile per organization, in one of five stages. A new agent starts at Supervised. A streak of correct decisions promotes it. A human override sets its progress back, and an error drops it one stage. The profile is per organization, so an agent earns its standing separately in every workspace.
Promotion thresholds
Promotion needs a run of consecutive correct decisions, and the count rises at each stage. The harness records a correct outcome on a clean run and an error outcome on a failure or an open circuit.
| Supervised to Assisted | 50 consecutive correct decisions |
| Assisted to Verified | 100 consecutive correct decisions |
| Verified to Trusted | 250 consecutive correct decisions |
| Trusted to Autonomous | 500 consecutive correct decisions |
Approval gates by stage
Each stage maps to a set of approval gates, defined in AUTONOMY_GATES. The further an agent has come, the fewer approvals it needs before it fills a form, submits it or takes a destructive action.
| Supervised | Fills: approval · Submits: approval · Destructive: approval |
| Assisted | Fills: auto · Submits: approval · Destructive: approval |
| Verified | Fills: auto · Submits: auto below a value threshold (default $10,000) · Destructive: approval |
| Trusted | Fills: auto · Submits: auto · Destructive: approval |
| Autonomous | Fills: auto · Submits: auto · Destructive: auto |
Autonomous
500 correct to reach
Fills auto · Submits auto · Destructive auto
Trusted
250 correct to reach
Fills auto · Submits auto · Destructive: approval
Verified
100 correct to reach
Fills auto · Submits auto below $10,000 · Destructive: approval
Assisted
50 correct to reach
Fills auto · Submits: approval · Destructive: approval
Supervised
New agents start here
Fills: approval · Submits: approval · Destructive: approval
Only the Autonomous stage acts destructively without a person, and an error is never waved through. In Reach, submit and destructive steps need approval at every stage.
auto_approve_below gate is set to 10,000. A submit under that value is approved automatically, and above it a person still signs off. Trusted and Autonomous have no value ceiling.What never relaxes
Trust changes how much an agent does without asking. Two rules hold at every stage:
- Destructive actions need approval at every stage except Autonomous, and in Reach a submit or destructive step needs approval even then.
- Destructive and error paths always go back to a person. The harness will drop a triage flag for a trusted agent only when the reason is neither destructive nor an error, and only when that stage's fill and submit gates are both already automatic.
How an action gets held
When an agent proposes an action, the harness checks the run against the agent's current gates. If the relevant gate (fill, submit or destructive) requires approval, or the proposal is above the Verified value threshold, the run is held as a triage item. The person reviewing it sees the proposed action, the gate that stopped it and the supporting context.
Example · A $12,400 payment held at Verified
An AP agent at the Verified stage prepares to submit a vendor payment of $12,400. Submits are automatic at Verified, but only below the $10,000 threshold. This one is above it, so the run is held in triage with the payment detail and the rule ("exceeds Verified auto-approve threshold"). A person approves, and the payment posts. The same agent submitting a $400 invoice would have gone ahead.
What is recorded
Every run is stored with the agent, the model used, the confidence score and the outcome. Every triage decision is stored with who resolved it. Audit-log retention follows the plan, from 7 days on Starter to 365 days on Enterprise. All of it is kept per organization.