The harness

An assembly line for software, staffed by AIs.

Every task moves through the same stations: research, plan, build, review. Which AI works each station is decided at run time — by what it’s proven it can do, what the rules allow, and how much capacity it has left this week. Even the lead can be swapped between phases. One supervisor keeps the whole floor moving and calls you only when a decision needs a human.

Read the studies Meet the players →

The route

The steps

Work moves station to station in a fixed order: Orchestra asks what you actually want, reads your code, drafts the plan two ways, stress-tests it, builds it, inspects it — and nothing leaves the floor until it passes final inspection. Step through the stations below.

The approach to a unit of work

step 01 / 13 · Unit of work

published seat studies

models trialled for seats

product fleets under one supervisor — Spindle, Flowmaster, ScriptGen

pull requests built, reviewed and merged

Glossary

The terms, in plain words.

unit of work: One scoped task, start to finish — deliberately large but still doable: it needs a real plan, touches many files, and takes multiple rounds of review. A rough description goes in; a finished, reviewed change comes out.
harness: The machinery around the AIs — the stations, checklists, and rules that turn many models into one production line.
agent: One AI model given a job, tools, and instructions. Several agents work on a task at different stages.
supervisor: The foreman. Never does the work itself; keeps every station moving and calls you when a decision needs a human.
seat: A station on the line — build this, inspect that. The job stays fixed; the model working it can change.
lead: The one model accountable for a phase of the work. One at a time — but which model that is can change between phases.
plan lock: The moment the plan is frozen. Nothing about the plan changes after this — and it's the cleanest point for the lead seat to change hands.
router: The rule that fills seats: qualified, allowed by policy, most headroom this week. No model picks another model.
pull request: A proposed change to the code, packaged so it can be reviewed — and accepted or rejected — as one unit.
merge: Accepting the finished change into the main codebase. The last step of every task.
gate: A pass/fail checklist run by the machinery, not by a model. If anything fails, the work goes back.
lint: Automatic code-style checking. One of the checks that must pass at the gate.
convergence: The point where review rounds stop finding new problems. Reviews continue until they get there.
pool: One subscription's worth of models from a provider — the Claude pool, the Codex pool, the GLM pool.
headroom: How much of a subscription's weekly allowance is left. Jobs go where the headroom is.
pin: Once a lead starts a phase, it keeps it. Reassignment only happens between phases, never mid-work.
context: An agent's working memory. It's finite — the supervisor swaps in a fresh lead before it fills up.

The seat contract

What is a seat?

A seat is a station on the line: build this part (build), inspect that one (plan.scrutiny), study the blueprints (discovery). The station and its rules stay the same even when the worker changes shifts — and workers do change. The router fills each seat from whichever models are qualified, allowed by policy, and have headroom this week. Even the lead seat for a phase is just another seat.

The seat ledger

Which model currently holds which job — a snapshot as of 2026-08-30, not a rule. Any seat can be re-filled from the roster.

policy filtered · see the players →

Lead seats

lead.*

lead.discovery

claude-opus-4-8

current · reassignable by phase

lead.plan

claude-opus-4-8

current · reassignable by phase

lead.plan-loop

claude-opus-4-8

flip pending · after one clean terra-led task

lead.build

gpt-5.6-terra

since 2026-08-29 · per lead-seat study

lead.review

gpt-5.6-terra

since 2026-08-29 · calls review done, then double-checks

lead.cleanup

gpt-5.6-terra

since 2026-08-25

Discovery seats

discovery.*

discovery.deep

claude-opus-4-8 ×3

reads the code in depth

discovery.breadth

claude-sonnet ×1

skims everything for breadth

discovery.lattice

glm-5.3 · 4–6 lanes

advises · never decides

discovery.web ×3

glm-5.3claude-opus-4-8gpt-5.6-terra

independent passes

discovery.invariant-map

claude-opus-4-8

works without seeing the others

Plan seats

plan.*

plan.co-draft ×2

claude-fable-5 · maxgpt-5.6-sol · deep

each drafts without seeing the other

plan.scrutiny

gpt-5.6-sol · seedgpt-5.6-terragpt-5.5

reviewers take turns

plan.loop-iterate

glm-5.3

quick trial pass

plan.consult

gpt-5.6-sol

tricky tasks only

Build seats

build.* / checkpoint.*

build.hard-step

claude-opus-4-8

the difficult steps

build.rote-step

claude-sonnet

+ automatic lint fixes

checkpoint.find

glm-5.3

checks every build step

checkpoint.confirm

gpt-5.6-terra

found by one, confirmed by another

Review seats

review.*

review.fresh-eyes

claude-sonnet · swarm

first pass · never saw the code

review.breadth

gpt-5.6-sol

wide-angle review

review.deep-rounds

gpt-5.6-terragpt-5.5

rounds continue until nothing new turns up

review.fill

gpt-5.3-codex-spark

max 4 at once

Gate + merge

gate.* / merge

gate.mechanical

no model seat

automatic pass/fail checklist

gate.clippy-floor

no model seat

code style checks must pass

gate.postgres-lane

no model seat

database changes must apply cleanly

merge

orchestrator

closes the pull request

Who holds each seat?

The seats stay put. The workers change shifts.

The router writes a pin every time it fills a seat: which seat, which model, when. This is that log for the five lead seats in the ledger above, on one fleet, from 2 Sept to 17 Sept: 62 seatings across 23 tasks. Build and review changed hands between two pools; plan was opus every time the router chose. Then on 14 Sept the owner benched a whole pool, and the same seats simply re-filled.

qualified ∩ allowed by policy ∩ headroom this week = eligible
coldest eligible pool → holds the seat until the phase ends

Lead seats · who actually sat

claude pool
codex pool
inherited

IIPlanlead.plan

claude-opus-4-8 11
inherited 4

claude-opus-4-8

IIIPlan looplead.plan-loop

claude-opus-4-8 1

claude-opus-4-8

IVBuildlead.build

claude-opus-4-8 6
gpt-5.6-terra 8

claude-opus-4-8

VReviewlead.review

claude-opus-4-8 11
gpt-5.6-terra 9

claude-opus-4-8

VIIICleanuplead.cleanup

claude-opus-4-8 1
gpt-5.6-terra 11

claude-opus-4-8

Lead seats · 62 pins

One square per pin, placed at the moment a lead was seated, from the router’s go-live on 2 Sept to 17 Sept. Pick a seat to read its pins, and what each pool had used that week at that moment.

claude pool
codex pool
inherited · router deferred to the running lead

Sep 04 · opus joins the lead roster
Sep 14 · codex pool benched

Latest pins · lead.build

showing the latest 5 of 14 pins · pools = weekly used %, claude · codex · glm

2026-09-14 18:50OUTclaude-opus-4-8ultracode59%70%39%re-seated at hand-off 13

2026-09-12 00:01WARM-XXXXXgpt-5.6-terraxhigh72%13%1%at phase start

2026-09-11 22:49W5gpt-5.6-terraxhigh70%11%0%re-seated at hand-off 9

2026-09-10 19:19INT7-XXXXXXXXXXXXXXXXXXclaude-opus-4-8ultracode88%66%69%re-seated at hand-off 4

2026-09-08 04:28W7-NAXXXXgpt-5.6-terraxhigh11%5%25%re-seated at hand-off 7

Phase IV · Build — checkpoint seats · 104 build steps

Every build step those leads ran stops at a checkpoint. One seat looks for problems, a different seat signs the step off, and heavier steps get more passes.

glm pool
codex pool

checkpoint.findlight steps

glm-5.3 17
gpt-5.6-terra 2

checkpoint.findstandard steps

glm-5.3 16
gpt-5.6-terra 16

checkpoint.findheavy · breadth pass

gpt-5.6-sol 53

checkpoint.findheavy · find pass a

glm-5.3 29
gpt-5.6-terra 24

checkpoint.findheavy · find pass b

gpt-5.6-terra 53

checkpoint.confirmevery step

terra + gpt-5.5, both sign 86
terra alone 18

104

Policy, visible in the log. glm-5.3 is barred from credential-tagged steps. On the 31 such steps it held the find seat 0 times and gpt-5.6-terra took all 31. Everywhere else GLM held 62 of 73.

Shift change · 2026-09-14

On 14 Sept the owner benched the whole codex pool. Nobody rewrote a seat. The roster behind each one changed, and the next task was seated from what was left.

IVBuildlead.build

gpt-5.6-terra
claude-opus-4-8

claude-opus-4-8

VReviewlead.review

gpt-5.6-terra

claude-opus-4-8

VIIICleanuplead.cleanup

gpt-5.6-terra

claude-opus-4-8

IIPlanplan.co-draft

gpt-5.6-sol

claude-fable-5-1

IIIPlan loopplan.scrutiny · seed

gpt-5.6-sol

claude-sonnet-5

IDiscoverydiscovery.web

gpt-5.6-terra

glm-5.3

VReviewreview.breadth

gpt-5.6-sol
glm-5.3 · swarm of 4

claude-sonnet-5
glm-5.3 · swarm of 4

IVBuildcheckpoint.confirm

gpt-5.6-terra
gpt-5.5

claude-sonnet-5

17 of 30 router seats re-filled · 0 renamed

Latest studies

every claim links to the raw data

07 · 23

Policy over time: how the seats were adjusted to the usage pools

Every recorded change to the outbound fleet's seat policy from 23 July to 15 September, placed on the throttling record that was being logged at that moment. Step through the changes and the pools move exactly as the readings say they did.

Record · 36 events

09 · 06

Can Fable 5.1 hold a codex review seat?

Eight frozen cells across three beds replayed Fable 5.1 at @high and @xhigh into the DECLARER, ITERATE, and FIND review seats — blind, on the exact diff the incumbent saw — against terra, gpt-5.5, and the GLM swarm. Fable is DECLARER/FIND-shaped: deep single-seam verify-after-fix, honest verdicts, zero fabricated cites in eight cells, and bugs terra missed. It is not ITERATE-shaped. @xhigh beat @high on every bed.

Ratified · declare + find

09 · 02

Can gpt-5.3-codex-spark hold a judgment seat?

We asked whether Spark could occupy a judgment seat in a frozen-cell benchmark. This run produced a narrower answer: the candidate exhausted its weekly capacity before the bench could say anything about capability.

Rejected · capacity

All 8 studies →

The practical result

More finished work from the models already on your bench.

View the players →Open Fleet Status →