The line
Waves and roles

The reference line of the AI Software Factory

Nine stations from idea to production

Every station shows who owns it: a person, an agent or code. No station trusts the one before it. Eight run today on the reference line; Release has its first real run at the launch planned for 1 November 2026. Scroll, or pick a box.

Start at station 1 Decide
79
tickets entered the lineMeasured
74
passed every automated gateMeasured
64
merged to mainMeasured
13
checks per run, medianMeasured
124
mutations proven across 40 runsMeasured
≈15 min
of a tech lead per merge by handEstimate
run log5 Oct 2026, replayed
iOverviewThe stations as boxes, the facts and the run log.
11 DecideA decision that is not in the log does not exist. A finding without a ticket is an opinion.

Station 1 of 9

Decide

A decision that is not in the log does not exist. A finding without a ticket is an opinion.

Person decides

Output
A decision record, then a spec or an architecture decision record.
From the logs
The log holds decisions back to April 2026. The ones that failed are marked FAILED and stay on the record.
1DecideApr 2026the decision log startsThe log holds decisions back to April 2026. The ones that failed are marked FAILED and stay on the record.Person decidesMeasured
22 TicketThe header is truth. The text is explanation.

Station 2 of 9

Ticket

The header is truth. The text is explanation.

Person decidesCode confirms

Output
A ticket file in git. Tickets for people live in the same place.
From the logs
226 tickets: 155 for agents, 71 for the human. 63 require a mutation proof. 34 carry a human gate.
2Ticket226tickets: 155 for agents, 71 for the human226 tickets: 155 for agents, 71 for the human. 63 require a mutation proof. 34 carry a human gate.Person decidesCode confirmsMeasured
33 OrchestrateA stale ticket stops before it starts. If a file the ticket expects is missing, it is blocked, not guessed.

Station 3 of 9

Orchestrate

A stale ticket stops before it starts. If a file the ticket expects is missing, it is blocked, not guessed.

Code confirms

Output
A prompt, an isolated workspace, a branch.
From the logs
The factory starts every hour from 07:00 to 01:00 and builds up to 6 tickets each time.
3Orchestrate6tickets per start at most, a start every hour from 07:00 to 01:00The factory starts every hour from 07:00 to 01:00 and builds up to 6 tickets each time.Code confirmsDesign value
44 BuildA block with a clear reason is a better result than a green suite over a widened scope.

Station 4 of 9

Build

A block with a clear reason is a better result than a green suite over a widened scope.

Agent makes or checks

Output
Changed files and a report: done or blocked, deviations, questions.
From the logs
94 runs: 79 passed, 15 stopped. 2,754 tool calls logged, 6 denied by the scope hook.
4Build79 / 94runs passed, 15 stopped with a reason94 runs: 79 passed, 15 stopped. 2,754 tool calls logged, 6 denied by the scope hook.Agent makes or checksMeasured
55 GateThe agent's claim that tests are green does not count. Only the orchestrator's run counts.

Station 5 of 9

Gate

The agent's claim that tests are green does not count. Only the orchestrator's run counts.

Code confirms

Output
A report with one row per gate.
From the logs
13 checks per run (median). 124 mutations proven across 40 runs. 76 of 79 passing runs needed one attempt.
5Gate13checks per run, median13 checks per run (median). 124 mutations proven across 40 runs. 76 of 79 passing runs needed one attempt.Code confirmsMeasured
66 ReviewGreen gates are necessary. They are not sufficient.

Station 6 of 9

Review

Green gates are necessary. They are not sufficient.

Agent makes or checksCode confirms

Output
A review file, a verdict, a comment on the pull request.
From the logs
72 tickets reviewed. In 47 the review took the merge decision away from the machine.
6Review47 / 72reviews took the merge decision from the machine72 tickets reviewed. In 47 the review took the merge decision away from the machine.Agent makes or checksCode confirmsUnconfirmed
77 MergeTests do not know whether a sentence in the terms of service is true.

Station 7 of 9

Merge

Tests do not know whether a sentence in the terms of service is true.

Person decidesCode confirms

Output
A merge commit on main.
From the logs
64 merged: 60 by a person, 4 by the factory. 10 wait for a person.
7Merge64merged: 60 by a person, 4 by the factory64 merged: 60 by a person, 4 by the factory. 10 wait for a person.Person decidesCode confirmsMeasured
88 ReleaseGates tell you the code works. Only a user tells you it was worth building.

Station 8 of 9 Designed, not yet run

Release

Gates tell you the code works. Only a user tells you it was worth building.

Person decides

Output
A release record: what went out, which commit.
From the logs
Not exercised yet. The checklist and the rollback procedure are written. The first real run is the e-šolstvo launch: GO or NO-GO on 29 October, launch planned for 1 November 2026.
8Release1 Nov 2026first real run, plannedNot exercised yet. The checklist and the rollback procedure are written. The first real run is the e-šolstvo launch: GO or NO-GO on 29 October, launch planned for 1 November 2026.Person decidesDesigned
99 ReportA model's claim is not a finding.

Station 9 of 9

Report

A model's claim is not a finding.

Code confirmsAgent makes or checks

Output
Run log, progress brief, memory episode, one to-do list.
From the logs
112 episodes, 67 active facts, 41 proposed facts waiting for a decision, 15 procedures.
9Report112memory episodes, 67 active facts112 episodes, 67 active facts, 41 proposed facts waiting for a decision, 15 procedures.Code confirmsAgent makes or checksMeasured
iBack to the overviewThat was the last station. Waves, roles and guardrails follow below.

Measured: the run log, ticket index and run reports of the reference line; 79, 74 and 64 cover 21 September to 5 October 2026. One product, one operator. The boxes replay the nine runs of 5 October 2026: ticket, tries, machine time and gate result come from the log, the other stations show their job. Release is designed, not yet run.

Overview. 9 stations follow.

Waves

The line grows in waves

A wave adds agents only after your line has met written targets at a checkpoint. Your sponsor decides there, on your numbers.

  1. Runs todayBuilt in your Pilot

    Base line

    8 agents

    Builds and checks real tickets

    The Pilot installs five of the eight on your team: ticket drafting, dev agent, review agent, status agent and morning brief, with gates for your stack and CI on your platform. The other three (an advisor session with memory, an outside watch on rules, an audit of the agents) follow in Implementation and in Run and Improve.

    Two people to start: a line owner and a decision owner with merge rights.

  2. Designed

    Wave 1

    13 agents · +5

    Contract before code

    Test author, release tester, ticket critic, merge card writer, stop analyst. Code checks the walls of the runner before every agent run.

    Target: a third less time per merge by hand.

  3. Designed

    Wave 2

    19 agents · +6

    Makers apart from checkers

    Repo scout, user tester, security reviewer, mission planner, release clerk, production watcher. Long work runs as a chain of short tickets.

    Target: 30% of merges automatic and stopped runs down to about 10%. A third person, an agent steward, owns the instructions.

  4. Designed

    Full line

    23 agents · +4

    More checks

    Decision council, screen reviewer, line mechanic, memory curator. They add checks, not speed, and each has to pass its own keep-or-kill test.

    At least three people for one team.

  5. Designed

    Production loop

    5 levels

    After the release

    A change is shown to users in steps, code watches it, a rule or a person decides, and every failure comes back as a ticket. From by hand to a closed loop.

    Sign-in, payments, database changes, legal text, anything a child sees and anything a revert cannot undo are never opened by a rule.

Waves 1 to 3 and the production loop are designs. The other agents are built on your processes and to your needs. Two of the ideas, blind judges and a mechanic, already run on the reference product's content line.

The people

Five roles. The line does the checking; people make the calls.

AI made writing code cheap and moved the queue to the people who check it: teams with high AI adoption merge 98% more pull requests, and review time grows 91% (Faros AI, 2025, outside source). The line is built the other way round. Code checks what code can check. A person sees a short card with the evidence and decides.

One rule on every card: a person decides, and never checks what code can check.

Architect

On the line: accepts the design · owns the rules · owns the knowledge base

Decides

The design note before the tickets: what the change is for, which module owns it, which decisions a revert cannot undo, and in which slices it ships. Which kind of change may later be handled by a rule.

What the line takes off their plate

Designed for wave 1: the ticket drafting session writes the first draft of all four parts of the design note before the architect joins, and code then checks on every change that the accepted design still holds.

30 min

per design note, 40 with security questions

Target

Tech lead

On the line: decision owner · release owner

Decides

Which tickets enter the line, what a stopped run needs, every merge the line may not make alone, and GO or NO-GO for a release.

What the line takes off their plate

Running the checks: 13 per build run (median, measured), a mutation proof where the ticket asks for one, and a separate review before anything reaches the merge. The merge card puts the evidence on one screen (designed for wave 1).

≈15 min

per merge by hand; at 8 hours a week, about 21 tickets a week at base line, 37 from wave 2

Estimate · planner

Developer

On the line: line owner (one developer) · agent steward (another, from wave 2) · owns the tickets that stay human

Decides

How the line runs day to day (as line owner), the examples agents should copy, the agents' written instructions, and the hard tickets that stay with a person.

What the line takes off their plate

Routine tickets. An agent stops and asks when a ticket cannot be finished inside its limits, instead of guessing, and the developer answers the question once, in the ticket.

155 / 226

tickets written for agents on the reference line. Keeping one agent's instructions healthy: about 1.5 hours a month (assumed).

Measured

QA

On the line: owns the release contract · builds the review drill · runs the user test

Decides

What must stay true after a release, which scenarios a release must pass, and which planted defects test the reviewers. The user test with real users before a public release.

What the line takes off their plate

Re-running every test on every run, and proving that the tests can fail: 124 mutations proven across 40 runs (measured). An agent that walks each release on staging and files evidence per scenario (designed for wave 1).

80% / 20%

of planted defects the review must catch / of clean changes it may block, at most

Target

Support

On the line: on call · merges the fix · rotates keys · platform and security contact

Decides

Revert, fix or hold. What agents may never read, run or reach. When a leaked key is rotated.

What the line takes off their plate

One log line per run and a morning brief of what waits for a person. Every defect found in production comes back as a ticket and a test that would have caught it. A weekly security scan drafts harden tickets (designed).

1 day

to rotate a leaked key, by rule

Design value

Also named in the Pilot plan: the sponsor who decides each next step on the numbers, the privacy contact, and the owner of the product metrics. One person may hold several roles. On the reference line one person holds all of them, and that is its known limit.

Guardrails

Six rules, and the incident behind each

None of these came from a framework. Each one is on the line because something went wrong first.

  1. A model's claim is not a finding.

    Incident

    An agent reviewing 79 curriculum files reported fixes it had never applied, then stopped half-way without saying so.

    What changed

    Deterministic verification. What code cannot confirm is labelled a judgement, never counted as a defect and never auto-fixed.

  2. Nobody grades their own work.

    Incident, 17 Sep 2026

    One agent wrote two files and reviewed them. They passed every gate built for them. 41% of the tasks were verbatim duplicates.

    What changed

    Two blind judges with different vantage points. Their disagreement is the signal and is never averaged.

  3. Test the instruments before the line runs.

    Incident, 17-18 Sep 2026

    Seven defects in detectors written by one author. One produced 17 false positives in 20 hits after it had driven 94 content edits.

    What changed

    An independent instrument tester, twice a day. A detector without a fresh test may not publish a number. A separate mechanic repairs it.

  4. A missing log line is a failure.

    Incident, 15-17 Sep 2026

    Scheduled runs reported success for three days and wrote nothing.

    What changed

    One run-log line per run, even on failure. A status agent looks for gaps every morning.

  5. A safeguard that never failed is not a safeguard.

    Finding, 21 Sep 2026

    On Windows a second shell tool was active by default. The deny rules and the scope hook did not apply to it.

    What changed

    A smoke test that makes a real model call and requires a recorded denial. It runs again after every tool upgrade.

  6. Contradictory output means stop and say so.

    Incident, 16 Sep 2026

    An output-filtering helper summarised a database command into a state that could not be true. The agent noticed and stopped.

    What changed

    A protocol rule, and state is read at the source. The helper was removed the same day.

What it runs on

Your stack gets its own gates

  • The reference line runs on a Next.js and PostgreSQL codebase on Windows, with CI on GitHub Actions.
  • Your stack needs its own gates; the Pilot builds them. The GitLab CI and Azure Pipelines templates have not run yet.
  • Model and tool accounts stay yours. Your code does not pass through mine.
  • The proof so far is one product and one operator. The line is installed on a new project built from scratch (0 to 1). Testing on larger codebases is under way.
Book a demo A 30-minute call. I open the real run logs on it.