The Human-Agent Operating Model
Define how people and agents divide decisions, execution, oversight, and accountability.
The chapter in one pass.
- Purpose: Define how people and agents divide decisions, execution, oversight, and accountability.
- Best for: Engineering, product, platform, and organizational leaders.
- Prerequisites: Operational Autonomy and Trust Calibration.
- Reading time: 20 minutes.
- You will learn: Which responsibilities remain human, which can be delegated, and how exception-first operation avoids approval theater.
Keep three ideas: humans own intent and material risk; agents perform bounded work and assemble evidence; and escalation should present a decision packet, not a transcript.
Reconstruct and defend this chapter’s architecture.
Reconstruct the architecture, name each boundary, and defend the tradeoffs.
Jump to validation criteriaOpen the source exercise
Draw the Product Owner, Mission Approver, Orchestrator, Worker, Validator, Engineering Lead, Risk Owner, and Release Approver. Connect them through the Mission lifecycle from intent to production verification.
For each boundary, state:
- the artifact transferred;
- the authority granted;
- the evidence required;
- the decision owner;
- the escalation condition; and
- what resumes after the decision.
Then redraw the model for a five-person startup and a regulated enterprise. Preserve the same accountability while changing how many people hold the roles.
1. The problem
Adding agents to an engineering organization does not define how the organization should work. Teams still need explicit decision rights, authority boundaries, handoffs, escalation rules, and accountability. Without them, humans either inspect every action or surrender control to systems that cannot own business risk.
The first failure produces approval theater. Operators become bottlenecks and approve routine work without understanding it. The second produces unsafe autonomy. Agents infer authority from prompts, validate their own output, and convert execution success into business acceptance.
An AI Software Factory needs an operating model that makes the division of responsibility unambiguous. The model must preserve human accountability while allowing agents to perform most bounded execution, iteration, validation, and evidence collection.
2. Why the problem exists
Traditional organizations embed authority in job titles, meetings, and tacit knowledge. A senior engineer knows which change needs security review. A product manager knows when scope has changed. A release manager understands which deployment requires an executive decision. Much of this context is never encoded in the delivery system.
Agents cannot safely inherit that ambiguity. They act through tools at machine speed, operate across shifts, and may continue when the person who initiated the work is unavailable. Their reasoning is probabilistic, their context is bounded, and their outputs can sound more certain than the evidence permits.
The organization must therefore convert tacit operating rules into explicit roles, policies, contracts, state transitions, and evidence requirements. This is organizational design expressed in software.
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:
| Decision | Accountable human owner |
|---|---|
| Mission outcome and priority | Product or Business Owner |
| Plan approval | Authorized Mission Approver |
| WorkOrder acceptance | Engineering Lead |
| Architecture exception | Principal Engineer or Architecture Owner |
| Security exception | Security Owner |
| Compliance exception | Compliance Owner |
| Material risk exception | Designated Risk Owner |
| Merge | Authorized code owner or Engineering Lead |
| Consequential production deployment | Release Approver defined by policy |
| Autonomy or learning promotion | Factory Governance Owner or Board |
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:
- Define. A human states the outcome, business reason, constraints, risk, sources of truth, and acceptance criteria.
- Research. An agent examines the repository and records citations, assumptions, and unknowns.
- Plan. The factory proposes a versioned execution and validation contract.
- Decide. A human approves, rejects, or requests revision of the exact Plan.
- Authorize. The factory materializes bounded WorkOrders and performs policy and capability preflight.
- Execute. Workers perform Tasks through immutable Attempts.
- Handoff. Each role records completed, incomplete, and unknown assertions, artifacts, risks, and the next owner.
- Validate. Independent validators evaluate frozen criteria.
- Recover. Failures produce new hypotheses and bounded corrective work.
- Accept. A human evaluates evidence, risk, deviations, and uncertainty.
- Release. Governed delivery proceeds through separate approval and production-verification states.
- 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.
4. Tradeoffs and alternatives
Strong role separation increases handoffs and can slow small changes. The answer is proportional control. Low-risk work may use lightweight automated handoffs while retaining the same lineage and evidence semantics.
Separate validators can repeat expensive work. That cost buys independence. Risk-based verifier selection, focused tests, artifact reuse with provenance, and sampling can control expense without letting the worker certify itself.
Human approval can become the throughput constraint. Removing approval is not the only remedy. Better intent, smaller WorkOrders, stronger evidence, clear recommendations, and exception-only routing often create more leverage with less risk.
Small organizations may combine human titles. One founder may be Product Owner, Engineering Lead, Mission Approver, and Release Approver. Technical separation between implementation and validation must remain. Separation of duties should increase with organizational scale and consequence.
5. Current Mission Control Implementation
This assessment uses Mission Control commit
8014d5af427b43ff5c5a63cfdf82ec92742c208c,
studied on 2026-08-08.
Mission Control's North Star explicitly assigns intent, judgment, governance, and approval to humans. Agents own bounded execution, iteration, validation, and evidence collection. The governed Mission contract defines four runtime roles: Orchestrator, Worker, Validator, and Operator.
Implemented mechanisms
Mission records retain state, owner, budget, stop condition, corrective limits, current Plan, active WorkOrder, blockers, and required human action. Plan submission freezes a proposed revision. Approval materializes linked assertions and WorkOrders idempotently while leaving dispatch as a separate decision.
Mission dispatch checks approved Plan authority, released WorkOrder state, predecessor handoff, budget, corrective limits, and serial mutation. Exactly one repository-mutating WorkOrder may be active in a Mission. Read-only work may run concurrently when the approved Plan permits it.
Worker and Validator handoffs record role, WorkOrder, WorkflowRun, assertion outcomes, commands, artifacts, risks, next action, and next owner. The handoff validator rejects overlapping assertion outcomes, false completeness, missing commands, or incomplete work without a stated risk.
Mission acceptance fails when assertions are missing, failed, stale, unvalidated, lack receipts, or use an unauthorized waiver. It also fails when WorkOrders or handoffs remain incomplete. Failed validation blocks the Mission and directs the operator toward bounded corrective work.
Capability assessment
| Operating capability | Status at studied commit | Interpretation |
|---|---|---|
| Human-defined Mission | Implemented | UI and Convex records capture governed intent and lifecycle. |
| Versioned Plan approval | Implemented | Submission freezes the proposal; approval releases exact WorkOrders and assertions. |
| Orchestrator, Worker, Validator, Operator roles | Implemented in the Mission contract | Role-specific permissions and state checks exist. |
| Structured handoffs | Implemented | Handoffs preserve assertion outcomes, commands, artifacts, risks, and next action. |
| Independent validation requirement | Implemented mechanism | Validator-run and receipt linkage are enforced for required assertions. |
| Corrective iteration | Implemented mechanism | Failed validation can block the Mission and bounded corrective work may be requested. |
| Exception-first operator experience | Product doctrine with partial UI proof | Required-human-action fields and decision surfaces exist; approval-fatigue reduction has not been measured. |
| Complete governance-role matrix | Partial | Runtime roles are clear, but all business, security, compliance, architecture, and release owners are not one enforced matrix. |
| Risk-proportional approval automation | Partial | Risk and approval controls exist; comprehensive cross-lifecycle policy proof remains incomplete. |
| Overnight autonomous operating shift | Product target, not proven here | Durable state exists, but this chapter did not demonstrate an unattended end-to-end shift. |
No fresh browser journey was performed for this chapter. The assessment is grounded in product contracts, source, and existing tests at the pinned commit. It does not promote the complete operating model into a proven V1 claim.
6. Future Vision
Mission Control should make decision rights visible at every state. Each Mission and WorkOrder should show who owns the next decision, why it is needed, the governing policy, deadline, evidence, options, and automatic resume behavior.
The operator should be able to begin an overnight shift by defining allowed Missions, risk boundaries, budgets, concurrency, notification rules, and stop conditions. The morning briefing should separate completed outcomes, review-ready changes, recovered failures, exhausted budgets, and decisions requiring human judgment.
Role compatibility should be checked before assignment. A worker must not also be the independent validator for the same material artifact. High-risk work should add security, architecture, compliance, or release owners through policy.
The operating model becomes proven only when repeated browser and runtime evidence shows that work survives process restart, handoff, validation failure, corrective execution, and delayed human review without bypassing authority.
7. Versioned references
- Mission Control North Star
- Mission Control V1 Product Strategy
- Governed Missions Contract
- Mission lifecycle commands
- Mission governance rules
- Mission execution reconciliation
- Mission detail operator view
- Mission planning workspace
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
- What does human-led, agent-executed engineering mean in practice?
- Which decisions must never belong to an agent?
- How does an Orchestrator differ from a Worker?
- Why can a Worker handoff not serve as validation evidence?
- What makes a handoff structurally complete?
- When should a factory interrupt a human?
- How do you prevent approval fatigue without removing accountability?
- Which roles may one person combine in a startup?
- What technical separation must remain in a small company?
- How do decision rights change from autonomy Level 2 to Level 3?
- What should happen after failed independent validation?
- How would you organize an overnight agent shift?
- Which operating-model capabilities does Mission Control currently prove?
- How would you measure whether the model reduces human cognitive load?
10. Whiteboard exercise
Draw the Product Owner, Mission Approver, Orchestrator, Worker, Validator, Engineering Lead, Risk Owner, and Release Approver. Connect them through the Mission lifecycle from intent to production verification.
For each boundary, state:
- the artifact transferred;
- the authority granted;
- the evidence required;
- the decision owner;
- the escalation condition; and
- what resumes after the decision.
Then redraw the model for a five-person startup and a regulated enterprise. Preserve the same accountability while changing how many people hold the roles.
11. Hands-on lab
Objective
Trace one Mission through every human and agent role, including failed validation, corrective work, and final acceptance.
Starting version
- Repository:
jaydubya818/MissionControl - Commit:
8014d5af427b43ff5c5a63cfdf82ec92742c208c - Study date: 2026-08-08
Tasks
- Create or select a governed Mission with explicit owner and criteria.
- Trace the Planner proposal and exact human-approved Plan revision.
- Identify each released Worker and Validator WorkOrder.
- Trace Orchestrator dispatch eligibility and serial mutation.
- Inspect the Worker handoff, including unknown and incomplete assertions.
- Cause or locate one Validator failure.
- Trace the operator attention item and bounded corrective iteration.
- Verify the new Worker Attempt does not erase the prior failure.
- Trace fresh Validator Evidence into Mission acceptance eligibility.
- Identify the human who makes the final acceptance decision.
Required evidence
- role and decision-rights matrix;
- approved Plan and released WorkOrder identities;
- Worker and Validator Attempt identities;
- complete handoff records;
- failure, escalation, and corrective-work records;
- criterion-linked independent Evidence;
- operator decision package; and
- teach-backs for a developer, engineering executive, and CEO.
Cleanup
Use seeded data and the controlled lab repository. Revoke temporary authority and preserve retained evidence without exposing credentials or sensitive paths.
Mastery standard
The chapter is mastered when I can assign every decision and action to the correct role, design a complete handoff, prevent unsafe self-approval, explain how autonomy changes human attention, and defend the operating model to a developer, CTO, CEO, and board without notes.
Review this chapter.
Challenge a claim, boundary, missing failure mode, unclear term, or unsupported evidence statement.
- Claim
- Boundary
- Failure
- Evidence