THE HARNESS MANUAL · Model Assignment
NOTES — ED. 2026.07
Not every task needs a giant.
Fit the model to the work, not your pride.
How to fill the last field in an agent file: model. Why every worker should not get the top-tier model, what the observed assignment table looks like, and this manual’s closing words.
FIG. 01CAST BY ROLE, NOT BY RANK — THE ENVOI
Model AssignmentWHY NOT ALL-MAX
Why Not Put the Best Model Everywhere?
MODELS · 2026The last field in the agent header: model. A beginner’s instinct says, “Give everyone the best.” Observation objects for three reasons. First, cost. Once a harness runs properly, the most frequent calls are not glamorous implementation but verification, comparison, and judgment. Put a top-tier model at every checkpoint and the safety bill exceeds the work itself. The industry’s empirical estimate is clear: mid-tier or lower models suffice for seven tenths of all work; concentrating the top tier on the remaining three tenths of critical tasks cuts the total cost by nearly half.
Second, speed. Gates must be fast. If the inspections at every commit and push are slow, people tire first, and tired people start turning inspections off. A bypassed safety device is worse than none: it creates the illusion that protection exists.
The third reason is the essential one: suitability for the role. Verification and judgment require discipline, not creativity. Follow the checklist in order, and say so when something differs from the standard. A top-tier model’s strength—reading between the lines and plausibly reinterpreting things—is a risk here, not a virtue. Once the judge starts reinterpreting the standard, the gate softens. We said the hand that creates and the hand that inspects must be different hands. Different hands use different muscles.
§ 01THE MEASURED CAST
The Observed Assignment Table
MODELS · 2026 Opening and counting every agent header across nine workshops reveals this assignment. Planners and implementers—roles that design specifications and build code in an open problem space—use top-tier models. Verifiers, simulator operators, content-consistency keepers, and cost accountants—roles that follow set procedures and fill comparison tables—use models one tier down. This two-tier arrangement is common to the later workshops.
The timeline is interesting. In the early workshops, all eight workers used top-tier models. The later a workshop was built, the more its roles were differentiated: the balance reached by the same person after experiencing both the bills and the quality of judgments. A lower bound is also observable. Two workshops explicitly wrote “lowest-tier models prohibited” into their constitutions. Even when procedures are mechanical, they do not drop below the point where judgment quality wavers. Another arrangement can be layered onto the timeline: a sandwich with expensive reasoning only at the two ends—planning and final verification—and lightweight implementation in the middle. Accurate planning makes implementation easier; accurate verification catches implementation mistakes. The assignment principle in one line: lower for more mechanical procedures, higher for more open judgment, never below the point where quality breaks.
The same principle applies to skills. The observed skill headers have an effort field: low for mechanical inspection procedures, high for procedures requiring judgment, such as completion-declaration gates and systematic debugging. Model tier and effort level are two dials for setting, role by role, how much attention to pay for on this task. That is assignment.
RoleTierReason
Planner · Implementer
Reason
An open problem space—the muscles of design and creationVerifier · Inspector
Reason
Enforcing a checklist—reinterpreting the standard is a risk, not a virtueSimulator operation · Consistency checks · Cost accounting
Reason
Set procedures and comparison tables—discipline is the coreCommon lower bound
Tier
Lowest tier prohibitedReason
No role goes below the point where judgment quality waversClosing WordsNOT A TECHNOLOGY OF DISTRUST
Closing Words—Not the Art of Distrust
ENVOI · 2026 If you have come this far, you will have noticed: this manual did not teach you how to hate machines. A harness is not the art of distrust; it is the art of building a structure that does not need trust. Paradoxically, only after that structure exists does the machine’s “I’m done” become believable. It now means the inspector has passed it, the sensor returned 0, and the gate opened.
Here is the minimal path once more, shortened for anyone who has only assigned work through chat. Write a sixty-line constitution. Create one inspector without editing tools. Create one completion-declaration gate skill. Lock the push checkpoint. These four things take one day this weekend—or half a day with the sample prompts in Part 03. From that day on, an exit code guards your evening, not the machine’s confession.
Tool behavior and settings vary by version. If the description differs from the actual screen, check your installed version and the official documentation, then run the sensors and gates yourself and compare the results.
— FIN —