How to build
a software factory?

Bennett Clement · FUNDA

Half of my work runs autonomously.

2,451 of 4,821 merged PRs since July.

Features: my sessions Operations: loops Benny, my ops bot

What is a software factory?

Diagram: The Pragmatic Engineer, "the agentic software factory inside OpenAI"

What most teams have today.

coding agent

CI

human review

Diagram: The Pragmatic Engineer, "the agentic software factory inside OpenAI" · annotations mine

How to get there.

01

Fix your codebase

02

Close the testing loop

03

Build loops

01

Fix your codebase.

01 · Codebase

Agents amplify your codebase.

Four layers of codebase guardrails: codebase, static analysis, prose, human review

Lauren Tan (@poteto), 2026-10-16 · x.com/poteto/status/2111133080526495985

01 · Codebase

Set up golden paths.

Four layers, with the codebase and static analysis layers highlighted

golden paths

architecture (DDD)

lints + CI

Lauren Tan (@poteto), 2026-10-16 · annotations mine

01 · Codebase

Rewrite. Code is (now) cheap.

Delete

Unused experiments
Unused features
Stale feature flags

Big refactorings

Unsafe type casts

1,482 → 0

in 4 days · rule → ratchet → drain

Core data model, 3 repos

~420 PRs · built in 3.5 weeks

Old code is code nobody uses: finished experiments, features with no users, flags that are always on.
Dead code is more code to maintain, for people and agents.
Big refactorings too, not only deletes. Add the lint, freeze a baseline, then let agents drain it in steps. Unsafe type casts: we froze 1,482 source sites, and about 80 PRs drained them to zero in 4 days. Now any new one fails the build.
Without the ratchet, new code refills the hole while you drain.
Migration: Large migrations used to cost too much, so teams lived with a bad model. Now the build takes weeks.
PR and file counts are upper bounds: some of them include other work.
The steps: put one API in front of the data → write to both old and new → move readers behind the API, behind a flag → observe in production → drop the old table.
Agents are more diligent than people: they take the backup and run the checks every time.
They also track what is done and what is not. The migration runs for weeks, but once the design doc exists, you only ask "what's the next step?"
The "someday" list is affordable now. The work was always worth doing; it just cost too much.

02

Close the testing loop.

02 · Testing

Make your agents see.

… so they can test like a human.

02 · Testing

Slopfork your dependencies.

(… for testing)

02 · Testing

Connect your agent to production.

Productionlive users

alerts / errors

Agentreads logs, traces

Pull requestthe fix

Humanreviews, merges

deploy

03

Build loops.

03 · Loops

Run gardening loops.

~2,200

PRs merged

~1,200

runs: nothing to do

03 · Loops

Build self-improving loops.

Learn from past fixes and trajectories.

03 · Loops

Take yourself out of the loop.

Always be thinking: can an agent do this? Can I trust it to?

Open challenges remain.

01

Always-on security agent

02

Safe production access for agents

03

Software factory infra

Code is cheap.
Rigor is not.

Tweet by Thariq (@trq212): very weird thing happening rn. Technical people are sad, they think their skills are now useless. Non-technical people are desperately wishing they were more technical so they could use agents better.

Thariq (@trq212), 2026-10-29

Thank you.

Connect on LinkedIn
linkedin.com/in/bennett-clement

How to builda software factory?

Half of my work runs autonomously.

What is a software factory?

What most teams have today.

How to get there.

Fix your codebase.

Agents amplify your codebase.

Set up golden paths.

Rewrite. Code is (now) cheap.

Close the testing loop.

Make your agents see.

Slopfork your dependencies.

Connect your agent to production.

Build loops.

Run gardening loops.

Build self-improving loops.

Take yourself out of the loop.

Open challenges remain.

Code is cheap.Rigor is not.

Thank you.

How to build
a software factory?

Code is cheap.
Rigor is not.