Part 1 · Operator’s Guide
The Bench
One accountable coordinator and bounded worker roles.
Public teaching edition · 1 October 2026
Who does what
One coordinator owns the task from the original request through verification, delivery and final status. Delegating a step does not transfer responsibility for the outcome.
| Role | Job | Boundary |
|---|---|---|
| Coordinator | Keep the ask, scope, owner, budget, evidence and next action clear | Cannot invent permission or suppress contrary evidence |
| Judgment worker | Resolve an ambiguous decision from evidence and constraints | May reopen the relevant source; must state uncertainty |
| Execution worker | Carry out a bounded step with the right tools | Cannot silently expand the task or spending limit |
| Reviewer | Check the artifact against the request and standard | Must inspect evidence rather than adopt the builder's conclusion |
| Human decision owner | Make decisions reserved for a person | Approval and quality review must be recorded truthfully |
One agent may fill several roles on a small job. Its own second pass is still a self-check. A fresh context is L2; a different model doing an independent review is L3. Keep required checks visible if unavailable.
Use a human-reserved baseline for Send, Publish, Spend, Merge and Hire. A narrower action may be delegated only by specific current authority and only when the applicable safety and tool rules permit it. Record the exact delegation; never infer it from a role or a login.
Record allowed actions, data boundary, budget, deadline, acceptance test and escalation owner. A claim, a login, urgency, confidence and a quiet human grant no new permission. Anyone may flag a material contradiction or stop the affected action.
Model agnostic by construction
Keep the request, rules, evidence, decisions, artifacts and permission boundaries outside any one model's memory. Ask whether a new worker could continue tomorrow from those records alone.
Use the least costly method that can pass the acceptance test: deterministic scripts for repeatable transformations; a smaller model for bounded bulk work; deeper reasoning for ambiguous judgments or expensive mistakes. Computer use is a tool requirement, not a fixed level of intelligence, and may coexist with any reasoning tier.
Set one total authorized budget before a multi-agent job. Include coordinator, workers, retries, tools and APIs, review, export and destination verification. Worker allowances come from that total. Reserve enough to check and close the result.
Track spent, remaining and in-flight work in the parent task. Unknown cost is not zero. If metering is incomplete, use bounded steps or time and stop before the ceiling. A fallback does not authorize a paid reload, a larger cap or a subscription. Compare cost per accepted result, including rework; historical token-price examples are not current quotes.