The Human-Agent Operating Model

3. Enduring Principle

Human-led, agent-executed engineering

Humans define valuable outcomes, grant authority, accept material risk, resolve ambiguity, and remain accountable. Agents research, plan, implement, test, validate, document, recover, and assemble evidence within granted boundaries.

Delegating execution does not delegate accountability. Human leadership moves upward from supervising activity to designing the system in which activity is safe and valuable.

Human decision rights

Humans retain decisions that define value or accept material consequences:

An agent may prepare the decision packet and make a recommendation. It cannot be the accountable owner.

Agent roles

Roles describe authority, not model personality.

Researcher or Planner investigates the repository and context, identifies unknowns, proposes a versioned Plan, and states assumptions. It cannot approve its own Plan.

Orchestrator sequences eligible WorkOrders, checks dependencies, dispatches within policy, tracks progress, requests corrective work, and escalates. It cannot widen scope, alter policy, or mark assertions passed.

Worker performs one authorized WorkOrder. It may implement, test, document, and produce artifacts. It cannot self-certify acceptance or start unrelated repository mutation.

Validator evaluates frozen criteria against exact artifacts through an independent execution path. It may report pass, fail, stale, unknown, or waiver-required. It cannot edit the implementation it certifies or approve its own waiver.

Recovery worker forms a new hypothesis from retained failure evidence and performs bounded corrective work. It does not erase the failed Attempt.

A simple workflow may combine several agent capabilities in one runtime. The logical roles and authority boundaries must remain explicit.

The operating cycle

The governed cycle is:

1. Define. A human states the outcome, business reason, constraints, risk, sources of truth, and acceptance criteria.
2. Research. An agent examines the repository and records citations, assumptions, and unknowns.
3. Plan. The factory proposes a versioned execution and validation contract.
4. Decide. A human approves, rejects, or requests revision of the exact Plan.
5. Authorize. The factory materializes bounded WorkOrders and performs policy and capability preflight.
6. Execute. Workers perform Tasks through immutable Attempts.
7. Handoff. Each role records completed, incomplete, and unknown assertions, artifacts, risks, and the next owner.
8. Validate. Independent validators evaluate frozen criteria.
9. Recover. Failures produce new hypotheses and bounded corrective work.
10. Accept. A human evaluates evidence, risk, deviations, and uncertainty.
11. Release. Governed delivery proceeds through separate approval and production-verification states.
12. Learn. The factory proposes reusable improvements. Humans promote them.

Handoffs replace conversational memory

A handoff must be a durable contract, not a chat summary. It records:

• producing and consuming roles;
• Mission, WorkOrder, and Attempt identities;
• completed, incomplete, and unknown criteria;
• commands, exit codes, artifacts, and changed files;
• known risks, blockers, assumptions, and uncertainty;
• next action and accountable owner; and
• whether the outcome is complete, incomplete, or needs human input.

Unknown is a valid state. Inventing continuity is not. The next role should not begin when the predecessor handoff is structurally incomplete.

Escalate judgment, not routine activity

The factory should interrupt a human when it lacks authority, evidence, a safe recovery path, or an unambiguous decision. Common escalation triggers include:

• conflicting or missing requirements;
• policy denial or expired approval;
• material scope or architecture change;
• validator disagreement;
• failed, stale, or missing evidence;
• security, privacy, legal, or compliance concern;
• exhausted budget or corrective-iteration limit;
• unexpected repository state or dependency;
• irreversible action; and
• uncertainty above the policy threshold.

An attention item must state the decision required, why autonomy stopped, affected scope, risk and urgency, available evidence, safe options, expected consequences, recommendation, uncertainty, and what resumes afterward.

Prevent approval fatigue

Approval volume should decrease as evidence quality and bounded autonomy improve. The factory should not ask humans to approve agent activity. It should ask them to decide about intent, exceptions, risk, and acceptance.

Routine low-risk work can proceed within policy. Surprises receive attention. High-risk work receives deeper review. The operator sees an evidence-backed decision packet rather than reconstructing events from chat and logs.

Autonomy changes the frequency, not the ownership, of decisions

At Level 1, humans initiate and review nearly every action. At Level 2, humans define WorkOrders and review all material outputs. At Level 3, agents may plan and execute while humans handle material risk and final accountability. At Level 4, policy may permit deployment for bounded low-risk classes. At Level 5, humans govern the factory system and its policies rather than routine work.

Human accountability remains at every level. Greater autonomy changes which decisions require individual intervention; it does not make the model the risk owner.

8. Notes and lessons learned

My current conclusions are:

• Human-led does not mean human-performed.
• Agent-executed does not mean agent-authorized.
• Roles should describe decision rights, not model personas.
• A durable handoff is part of the product contract, not optional reporting.
• Unknown information must remain unknown until resolved.
• Approval quality matters more than approval count.
• The factory should escalate surprises and judgment, not routine activity.
• Failed validation is a normal feedback path, not proof that the operating model failed.
• Small companies may combine people, but they must preserve technical independence.
• Humans govern policy and risk even when they stop reviewing every action.

9. Interview and discussion questions

1. What does human-led, agent-executed engineering mean in practice?
2. Which decisions must never belong to an agent?
3. How does an Orchestrator differ from a Worker?
4. Why can a Worker handoff not serve as validation evidence?
5. What makes a handoff structurally complete?
6. When should a factory interrupt a human?
7. How do you prevent approval fatigue without removing accountability?
8. Which roles may one person combine in a startup?
9. What technical separation must remain in a small company?
10. How do decision rights change from autonomy Level 2 to Level 3?
11. What should happen after failed independent validation?
12. How would you organize an overnight agent shift?
13. Which operating-model capabilities does Mission Control currently prove?
14. How would you measure whether the model reduces human cognitive load?