Meetup · 20 minutes

The agentic software factory

It's the 2022 one with a single box swapped. And that moves everything else.

Brahim Bousnguar · tech lead & e-commerce integration · heybrahim.com

EN · FR

The point

Coding agents didn't invent the software factory. They swapped one box in it.

In 2022

A person builds the feature. Hours or days.

Today

An agent builds it. Minutes or hours.

Everything else stayed. And the slow part of shipping moved from building to review.

Where the picture comes from

Dex Horthy, keynote State of the Software Factory, AGNTCon + MCPCon Europe, September 2026.

  • The idea of a software factory goes back to 1968, the year the term "software engineering" was coined.
  • He draws the factory most teams ran in 2022.
  • I liked that diagram so much I redrew it.

Video: youtube.com/watch?v=-w6RDNBAI0E, at 43:53.

Stage 1

The 2022 factory already had everything

  • A team, a tracker, someone who builds.
  • A pull request, CI, a human who tests and reviews.
  • Prod, users, monitoring.
  • Two loops: rework and feedback.

None of this is AI. It's how most teams have shipped for a decade.

The software factory, from 2022 to lights-off A vertical flow: team, stuff to do, build, pull request, checks, review, prod, then users and monitoring. A rework loop runs from review back to build, and a feedback loop runs from users and monitoring back to the team. In 2022 a person builds and people review. In the agentic factory an agent stack builds and agentic review joins the human one. In the lights-off factory the human review is struck out. LOOP 1 · REWORK TEAM CEO / vision · product managers · engineers STUFF TO DO Linear · Jira · GitHub Issues BUILD HOURS–DAYS MINUTES–HOURS someone builds the thing automated testing manual testing, e.g. in a browser an agent builds the thing orchestration harness sandbox model automated + agentic testing (computer use) PULL REQUEST CHECKS ci/cd · unit tests static scanning · security scans REVIEW HOURS–DAYS HOURS–DAYS · BOTTLENECK SKIPPED a human tests the change a human reviews the change agentic code review agentic regression tests a human tests the change a human reviews the change NO THANKS PROD rollout · feature flags · canaries USERS use it, complain MONITORING pages you at 3 a.m.

Stage 2

Swap the builder

  • Build becomes a stack: orchestration, harness, sandbox, model.
  • Agents test their own work, sometimes in a real browser.
  • Build: minutes to hours. Review: still hours to days.

Teams add agentic code review and agentic regression tests. It helps, but review often stays the bottleneck.

The software factory, from 2022 to lights-off A vertical flow: team, stuff to do, build, pull request, checks, review, prod, then users and monitoring. A rework loop runs from review back to build, and a feedback loop runs from users and monitoring back to the team. In 2022 a person builds and people review. In the agentic factory an agent stack builds and agentic review joins the human one. In the lights-off factory the human review is struck out. LOOP 1 · REWORK TEAM CEO / vision · product managers · engineers STUFF TO DO Linear · Jira · GitHub Issues BUILD HOURS–DAYS MINUTES–HOURS someone builds the thing automated testing manual testing, e.g. in a browser an agent builds the thing orchestration harness sandbox model automated + agentic testing (computer use) PULL REQUEST CHECKS ci/cd · unit tests static scanning · security scans REVIEW HOURS–DAYS HOURS–DAYS · BOTTLENECK SKIPPED a human tests the change a human reviews the change agentic code review agentic regression tests a human tests the change a human reviews the change NO THANKS PROD rollout · feature flags · canaries USERS use it, complain MONITORING pages you at 3 a.m.

The bottleneck moves

min – hto build,
with an agent
h – daysto review,
still

Building faster without touching review just fills a queue.

Stage 3

The lights-off factory

  • Nobody reviews: the checks decide.
  • With enough CI, agentic testing and monitoring, it could almost work.
  • The rework loop disappears.
The software factory, from 2022 to lights-off A vertical flow: team, stuff to do, build, pull request, checks, review, prod, then users and monitoring. A rework loop runs from review back to build, and a feedback loop runs from users and monitoring back to the team. In 2022 a person builds and people review. In the agentic factory an agent stack builds and agentic review joins the human one. In the lights-off factory the human review is struck out. LOOP 1 · REWORK TEAM CEO / vision · product managers · engineers STUFF TO DO Linear · Jira · GitHub Issues BUILD HOURS–DAYS MINUTES–HOURS someone builds the thing automated testing manual testing, e.g. in a browser an agent builds the thing orchestration harness sandbox model automated + agentic testing (computer use) PULL REQUEST CHECKS ci/cd · unit tests static scanning · security scans REVIEW HOURS–DAYS HOURS–DAYS · BOTTLENECK SKIPPED a human tests the change a human reviews the change agentic code review agentic regression tests a human tests the change a human reviews the change NO THANKS PROD rollout · feature flags · canaries USERS use it, complain MONITORING pages you at 3 a.m.

The catch: maintainability

What a test can tell you

This feature works.

What nothing tells you quickly

Will this code still be easy to change in two months?

Models get better at what a verifier can score. Maintainability has no fast verifier.

What SlopCodeBench measures

University of Wisconsin–Madison: agents don't get the whole spec up front. They get it in checkpoints and have to extend their own code, the way real projects grow.

14.8%the best model at launch
(GPT 5.5, as quoted by Dex)
~95%on single-shot benchmarks
like Terminal-Bench, per Dex

Dex's conclusion

For now, read the code.

If unattended code gets worse over time, human review isn't a cost to cut. It's the quality control.

What actually changed

Question2022Today
Who builds?A personAn agent stack you configure
Where does the time go?Typing codePlanning it, reviewing it, wiring up feedback
The job?WritingDeciding

Five practices that hold up

  1. Plan before you build. Twenty or thirty minutes up front saves hours at review.
  2. Give agents back pressure. A browser, curl, the running app, so they check their work before a human does.
  3. Feed loop 2 straight in. A pager alert that arrives with the PR the agent thinks fixes it.
  4. Let agents experiment. Feature flags, three versions of a UI, and real users pick.
  5. Compound. Read past agent sessions and turn what went wrong into skills and memory.

These practices are from Dex's talk.

My factory runs on stage 2

Issuefirst, always
Agentits own branch
Prooftests, a real browser
2nd opiniona model from another vendor
My PRI read it, I merge it

The slow part is my review queue. Exactly as drawn.

Guardrails that came from real mistakes

Check what was produced

A text-to-speech model invented sentences in the audio of my notes. Every chunk is now transcribed with Whisper and re-recorded if it adds or drops a word.

An empty review is not a pass

Reviewer models sometimes come back with nothing. The pipeline says so in the first line instead of calling the change clean.

No invented facts

No figure, client or experience goes on my site unless I gave it.

On the integration side

An agent that can't see the real data guesses. On my MuleSoft mission, I first wrote an MCP server that reads the Anypoint logs.

  • The public, open-source version: mulewatch, read-only.
  • Same idea for my portfolio: api.heybrahim.com/mcp.

The back pressure of practice 2, applied to integration.

Thank you

The full note, with the animated diagram: heybrahim.com/notes/software-factory-2022-to-lights-off.html

heybrahim.com · b.bousnguar@gmail.com · Dex's keynote : youtube.com/watch?v=-w6RDNBAI0E

EN · FR