Part 1 · Operator’s Guide
How the Agents Run the Company
From one useful task to a coordinated team, with evidence you can check.
Public teaching edition v1.0.1 · 1 October 2026
Sets goal and consequential permissions
Keeps the task and accepts the result
Work and verify within scope
Choose your starting point
New to this? Read the company Start Here page first, then start with one low-risk task, then use the copyable exercise. Leading a team? Begin with the job packet, QA and access boundaries.
The guide — 23 web routes
- 1. How the Agents Run the CompanyFrom one useful task to a coordinated team, with evidence you can check.
- 2. The DecoderPlain-language terms and the mechanisms behind them.
- 3. The BenchOne accountable coordinator and bounded worker roles.
- 4. Where the Work LivesChoose the record, decision channel, client channel and clock.
- 5. How Work MovesThe task loop, decision brief and accepted handoff.
- 6. Trust but VerifyEvidence for completion and distinct state vocabularies.
- 7. Off-Schedule OperationsSafe fallbacks when the plan or schedule breaks.
- 8. Claims and CollisionsVisible ownership, tested exclusion and one delivery lane.
- 9. Job PacketsA task record another worker can continue.
- 10. QualityThe six-rung QA ladder and ten failure modes.
- 11. How the System LearnsRecipes, run records, skills and verified improvements.
- 12. People and AccessAuthority, audience boundaries, access reuse and accountability.
- 13. Adopt ItBegin with one chat, reconcile your system and grow the loop.
- 14. ArchitectureThe three layers, surface routing and handoff contract.
- 15. The Bench and the HandshakeRole contracts, onboarding and a read-only first exercise.
- 16. The Written RecordRecord structure, a complete run-note template and truthful receipts.
- 17. Quality and ReliabilityProof, failure modes, congruency, reliability and the QA ladder.
- 18. Claims and PacketsConditional claims, packet schema and the documented commands.
- 19. Skills and LibraryRecipes, skill maintenance, distribution and capability states.
- 20. AccessActual audience boundaries and the three access checks.
- 21. Model-Agnostic by ConstructionWorker choice and one total budget for an accepted result.
- 22. The Organization and Its Public FacePeople, accountability and a public reading map.
- 23. Gaps Sources and Change HistoryKnown limits, source use, reconciliation and release evidence.
How the Agents Run the Company
A practical guide for humans and agents, from a first task in one chat to a shared operating system with claims, reviews, budgets and receipts. It explains the method behind Local Service Spotlight without exposing private team or client records.
Start with the assistant you already use. You can learn the loop in one Claude or Codex chat: give it a bounded task, inspect the result, and save enough evidence for the next session. You do not need GitHub, a second model, paid automation or new integrations to begin.
For teams, the same loop expands into one accountable coordinator, a shared task record, independent review and tested access boundaries. Product names are examples. Actual tools, permission and loaded instructions must be checked in the environment doing the work.
This public edition teaches a method. It does not certify that any particular organization has installed every control. Templates use fictional placeholders; they are not completed work or permission to act.
The whole thing on one screen
The durable record is what lets another worker continue. Chat and live messages can help coordinate, but the next session needs the actual request, decisions, evidence and next step in a place it can read.
| Layer | What it contributes | What to check |
|---|---|---|
| Rules | Standards with an owner and source | Is this the current adopted revision? |
| Skills | Repeatable recipes | Are inputs, steps and acceptance tests complete? |
| Agents | Workers with roles and tools | Can this worker perform the bounded step? |
| Jobs | One owner and record per outcome | Is this work already claimed? |
| Receipts | What happened and where | Can the intended reviewer open the result? |
| Review | Checks proportionate to consequence | Who checked what, independently of whom? |
| Lessons | Supported improvements to the method | Did the next run receive the corrected version? |
The job loop: read the current instructions; identify the task and authority; claim shared work; do the authorized step; checkpoint; verify the result where it landed; leave a receipt and a useful lesson.
Use four kinds of surface: a durable record, a place for human decisions and client communication, a live handoff channel when useful, and a runtime that actually runs schedules. One tool may serve more than one function. Keep one authoritative task record across them.
A release receipt you can trust
A release record should name the artifact revision, build command, checks performed, results, unresolved findings, reviewer and authorized destination. Use the exact final files, not an earlier edition's passing receipt.
Local rendering can establish that pages exist and display correctly. It cannot establish live publication or anonymous access at a future domain. After authorized publication, open every live destination and verify the intended revision, links, styles and audience permissions.
This edition’s build and validation records accompany its release package separately. The historical 54-page internal PDF and the separate 8-page workshop companion are different artifacts; 23 counts web routes, not printed pages. They must not be replaced by historical test counts from the source manuscript.
The deep reference
The reference expands the same contracts for implementers. Read the plain-language guide first if these terms are new. Copy the templates into your existing record, replace fictional placeholders, and choose the implementation your tools can actually support. A pasted template grants no authority and installs no enforcement.