The AI Software Factory — agents that ship.

An organizational system where autonomous agents run the software development lifecycle: a specification triggers generation, testing, deployment. The agents iterate until everything passes — you approve the PR.

§ 01 / Premise

A simple assistant produces fragile, untested code.

  1. 01The factory starts with a set of specifications,
  2. 02writes the tests and plans quality assurance before any code is produced,
  3. 03iterates until everything passes green,
  4. 04cross-checks every change,
  5. 05and feeds a process of continuous improvement and operational excellence.

The backlog overflows, fixes wait, small requests pile up behind big ones. The factory picks up intent where it is born — an issue, a support ticket, an alert — and carries it all the way to a tested PR. And nothing moves through it uninstrumented: the definition of done is written before the first line of code, every gate blocks until it is cleared, and every step leaves its measure — lead time, iterations, cost.

  • 01 — Agentic workflows

    Specialized agents, working as a crew.

    Planning, front end, testing, verification: each agent holds a station. Work flows from one to the next like a production line — each doing what it does best.

  • 02 — Convergence

    Iterating beats guessing.

    The success criteria are pre-registered — scenarios, tests, quality thresholds — then the system reworks the piece until everything passes, with no human intervention between attempts.

  • 03 — Dark factory

    Humans monitor, approve, merge.

    Generation, testing and deployment run on their own inside the workshop. Your role shifts to what matters: reading the signals, deciding, signing the merge.

§ 02 / The line

From intent to PR, in a closed loop.

Every piece follows the same route: intent becomes a specification, the specification becomes tested code. When a test fails, the line loops back — as many times as it takes.

Convergence line — schematic viewThe loop — iterate to green
In: an intentOut: an approved PR

Fig. 01 — The line loops through generation and tests until convergence

  • INTENTissue · ticket · alert
  • SPECscenarios · constraints
  • GENERATIONcoding agents
  • TESTSvalidation harness
  • CONVERGENCEconstraints: 12/12 ✓
  • REVIEW — YOURSPR #482 · ready
  • AGT·PLAN

    Planning

    takes: the intent · yields: the spec

  • AGT·CODE

    Generation

    takes: the spec · yields: the code

  • AGT·TEST

    Testing

    takes: the code · yields: the verdicts

  • AGT·VERIF

    Verification

    takes: the whole · yields: the PR

§ 03 / What it takes on

What the factory takes on.

Four families of work, from report to PR. The numbers are illustrative — yours get measured on your pipeline.

  • Bug

    From report to fix, no queue.

    A bug filed in Azure DevOps — or caught by a monitoring alert — triggers the line: reproduce it as a test, fix it, verify nothing regressed. The PR arrives with the test that proves the fix.

    issue → spec → 2 iterations → PR
    0 hfrom report to review-ready PRTypical scenario
  • Feature

    Intent becomes a spec, the spec becomes code.

    A request comes in through Jira or Linear; the planning agent turns it into scenarios and constraints, which you validate in Teams before the workshop starts.

    request → validated spec → 3 iterations → PR
    ×0delivery cadenceTypical scenario
  • Migration

    Large repetitive jobs, finally affordable.

    Version upgrades, framework changes, porting a module: the same constraints applied a hundred times. This is where convergence shines — every batch passes through the same tests.

    plan → batches → n iterations → PR per batch
    −0 %migration effortTypical scenario
  • Technical debt

    Refactoring as a background task.

    The factory works the debt continuously: every PR is measured against your code-quality metrics, and the gains compound while the team stays on value.

    tooled report → spec → PR
    +0 ptscode-quality scoreTypical scenario
§ 04 / Hookup

The factory plugs into your world.

The factory embraces your environment: your management tools, your repos, your cloud, the model your governance approves. Nothing else to adopt.

  • 01

    Your management tools

    It connects to any project-management tool — work enters and leaves where your teams already live.

  • 02

    Measurement included

    DORA and code-quality metrics ship with the factory — and every step logs its lead time, its iterations and its cost. You see cadence, stability and code health, continuously.

  • 03

    Your environment

    Cloud or on-premises, monorepo or a hundred repos: the factory adapts to your environment, not the other way around.

  • 04

    Your model

    Any large-model provider — the one your governance approves, swappable without touching the line.

Hookup sheet
  • Project-management toolsyours
  • Code repositoriesyours
  • Team channelsyours
  • AI modelthe one you choose
  • Environmentyours — cloud or on-premises
  • Measurement & metricsincluded

Nothing to replace — the factory connects to what is already in place.

§ 05 / Control room

The factory runs. You watch the signals.

Nobody reads the code line by line — everybody sees what matters. DORA metrics, iterations, cost per step: measurement is a mirror for you, never an objective handed to the agents.

Control room — 1 pipeline · 4 agents
  • 0 hDORA · Elite

    from intent to review-ready PR — lead time for changes

  • ×0DORA · Elite

    deployment frequency

  • −0 %DORA · High

    change failure rate

  • 0 minDORA · Elite

    to restore service after an incident

Iterations to convergence

1234+
median: 2 iterationsiterations
Log — latest piece
  • 13:42 · intent received · issue #517 — Azure DevOps
  • 13:44 · spec generated · 3 scenarios · 12 constraints
  • 13:51 · iteration 02 · 2 red tests → rework
  • 14:02 · tests green · 12/12 constraints met
  • 14:03 · PR #482 ready for review · assigned to m.gagnon
  • 14:17 · approved and merged · deploying
Agents
  • AGT·PLANstanding by
  • AGT·CODEgenerating · PR #483
  • AGT·TESTrunning · 214 tests
  • AGT·VERIFqueued

Calibration — estimated vs observed

estimated range p50–p80observed median

the gap feeds the next cycle’s optimism correction

Human attention vs agent work

  • Bug1 h · 6 h
  • Feature2 h · 14 h
  • Migration3 h · 26 h
  • Technical debt1 h · 8 h
human attention — the scarce resourceagent work

1 h of human attention for 8 h of agent work

Typical scenario — your numbers get measured on your pipeline, within the first weeks

§ 06 / The review

Autonomous in the workshop, human at the gate.

Before reaching the gate, every change faces adversarial review: independent verifiers try to refute each finding, and only what survives makes it to you — proof attached.

  • 01

    A single exit door

    Every change leaves the factory as a PR, and that PR waits for your signature. Approving, merging: that gesture stays human.

  • 02

    Your constraints first

    Scenarios, quality thresholds, performance budgets: you set the rules before the first line. The factory converges toward your constraints — never its own.

  • 03

    Everything replays

    Every iteration, every test, every agent decision is logged. You can trace the thread of any PR, months later.

Review sheet — PR #482
  • constraints12/12 ✓
  • tests214/214 green ✓
  • coverage96 %
  • scenarios3/3 validated ✓
  • iterations2
  • signature requiredyours

Multiply your development velocity.