Every box is a seat that a human or an agent can fill. Each leader gets a share of capacity from above and splits it among the seats below. Leaders can hire agents onto probation, bench the ones that underperform, and rehire them later. Every agent seat has a human overseer who can see it, pause it and undo its work.
A seat holds the authority, the budget share and the reports. Who fills it is one field, so handing a seat to an agent or taking it back changes nothing else. Agent seats carry an autonomy level that rises as the scorecard earns it, and every agent seat has a named human overseer with a backup.
A human fills the seat and makes every decision. Agents only do assigned tasks.
The agent proposes hires, benches and rehires. The overseer approves each one.
The agent acts within its limits. The overseer reviews afterwards and can undo.
The agent acts within its limits. Only exceptions reach a human.
Capacity is the money, tokens and concurrent job slots a unit may use. Each leader gets a share of its parent's capacity and splits it among its reports. Whatever a leader doesn't assign stays in its reserve. Here is one monthly budget traced from the board to a single worker agent.
This is the Platform Lead's seat. Drag a slider to change an agent's share; the total can't go past 100%. Benching frees a share, and rehired or new agents start on probation at 0%. Change the autonomy level to see who acts: at L1 a human runs the seat, at L2 the lead agent's hires, benches and rehires wait for the overseer, and at L3 and L4 it acts and reports. Budget splits apply right away at every level and go in the ledger.
Every agent seat has a named human overseer. The control room shows each seat's status, spend and alerts, with a pause button on every card. Pause everything is plain software, so stopping work never depends on an agent cooperating. Pausing the Platform Lead here also pauses the console above.
Daily digest, and 2% of work sampled for human review.
Alerts, 10% sampled, and every change ready to roll back.
A human approves before anything takes effect.
Agent definitions live in the registry. A lead picks one that its authority allows.
Shadow mode at 0% capacity. Its outputs are compared with the crew's, and nothing it does takes effect.
Gets a real share of capacity and its work takes effect, within its permits and checklists.
HR Agent scores it on a schedule: success rate, cost per job, checklist pass rate, human overrides.
Lead raises its share, or the executive makes it a crew lead.
A change set updates its definition, then it goes back to probation.
Capacity drops to 0 and access is revoked. Its history stays, so it can be rehired.
Needs a human's approval. The ledger keeps its record.
| Level | Capacity | Hire and fire | Approves change sets | Escalates to |
|---|---|---|---|---|
| Board (humans) | Sets the total budget | The CEO seat, and every retirement | All, final say | — |
| Overseer (human) | Watches spend in its area | Can take any seat back from an agent | Approves moves at L2; can roll back any change | Board |
| CEO | Splits across divisions, holds the reserve | Executives, with board approval | High risk | Board |
| Executive | Splits across its crews | Crew leads; proposes retirements | Medium risk, own division | CEO |
| Crew lead | Splits across its agents | Hire to probation, bench, rehire | Low risk, own crew | Executive |
| Worker agent | Uses its own share | None | Proposes only | Crew lead |
| Staff functions | Enforce caps, don't allocate | HR runs reviews and recommends | Policy checks on every change set | CEO and Board |
seat: platform-lead occupant: { type: agent, agent: lead-agent@v2 } autonomy: L2 # suggest; overseer approves overseer: { user: priya, backup: sam } reports_to: cto capacity: from_parent: 60% # of the CTO's share allocations: coder: 40% tester: 35% reviewer: 25% authority: can_hire: [catalog:engineering/*] # into probation only can_bench: true can_retire: false # a human approves max_agents: 6 approve_change_sets: { max_risk: low } review: every: 7d scorecard: [success_rate, cost_per_job, checklist_pass, human_overrides] bench_if: { score_below: 60, reviews: 2 } probation: { mode: shadow, min_jobs: 20, promote_if: "agreement >= 85%" }