Skip to content

Autonomy matrix

Orqane classifies every tool call into an impact class, then checks it against the employee’s autonomy level. This page shows the result for every combination. For how levels work and how employees get promoted, see Autonomy levels.

Value Meaning
Blocked The employee can’t perform the action. It can only describe it as a suggestion.
Approval The employee prepares the action, creates an approval and pauses the run until a human answers.
Threshold The employee acts on its own below the thresholds set on its Autonomy tab, and asks for approval above them.
Allowed The employee acts on its own. The action is recorded in Activity.
Allowed + report The employee acts on its own and reports the action to its human manager in its run summary and daily digest.
Impact class L0 Suggest L1 Approve L2 Act within limits L3 Autonomous L4 Delegate
Read Allowed Allowed Allowed Allowed Allowed
Draft Allowed Allowed Allowed Allowed Allowed
Write internal Blocked Approval Threshold Allowed + report Allowed + report
Communicate Blocked Approval Threshold Allowed + report Allowed + report
Spend Blocked Approval Threshold Allowed + report Allowed + report
Delegate Blocked Blocked Blocked Blocked Allowed + report
Irreversible Blocked Approval Approval Approval Approval
Impact class Examples Default thresholds at L2
Read Search the web, read a file, look up a CRM record, query a database None
Draft Write a document, prepare an email, save a note to memory None
Write internal Create a HubSpot deal, update a ticket, add a row to a sheet Records changed per run, allowed hours
Communicate Send an email, reply to a customer, post in a shared Slack channel Recipients per message, allowed domains, allowed hours
Spend Pay an API with x402, buy a dataset, pay another agent Spend per payment, spend per task
Delegate Create a sub-task for another employee, hire a marketplace agent, send a request to an outside agent over A2A Delegation budget (L4 only)
Irreversible Delete records, sign a document, publish on a public website or social account None, always asks

Orqane checks every action in this order. The first rule that blocks or requires an approval wins.

  1. Guardrails. Unsafe content or actions are blocked on every step, at every level.
  2. Workspace policies. A policy that blocks an action blocks it for every employee. A policy that requires an approval applies even at L3 and L4. See Policies.
  3. Budget and wallet limits. A payment above a spend limit, or to a merchant outside the allowlist, is blocked whatever the level. When the budget is used up, any further spend needs an approval by default.
  4. Autonomy level. The matrix above.
  5. Thresholds. At L2, the values on the employee’s Autonomy tab.

A policy can only make the matrix stricter, with one exception: a policy can explicitly allow an Irreversible action for a named employee or team, such as “Support may close tickets”. Only an Admin can create that kind of policy.

  • An Irreversible action requires an approval at every level, unless a policy explicitly allows it.
  • The approval shows what will be lost or made public, and can’t be answered by the employee that requested it, or by another employee.
  • Approvals for Irreversible actions can’t be granted in bulk. Each one is answered on its own.
  • After approval, Orqane records the action and the approver in the audit log.
  • Hand-offs. Passing a task to a human, or passing a whole task to another employee that accepts hand-offs, is allowed at every level. Splitting work into sub-tasks for other employees, as a manager employee does, is Delegate, so manager employees run at L4. Each sub-task runs under the level, policies and budget of the employee that receives it.
  • Messages inside the team. Messages between an employee and the members of its own team count as Draft. Messages to anyone else count as Communicate.
  • Tests. In a sandbox test, Read and Draft actions run for real. Every other action runs as a dry run.
  • Callers over MCP and A2A. An employee called over MCP or A2A keeps the same level, thresholds and policies.