AI CompendiumA public chart of the language
Home / Patterns of Work / Software Factory
Patterns of Work

Software Factory

A pipeline where agents take tickets and produce code, with checks along the way.

Definition

A workflow that treats software delivery as a production line: specs in, agents implement, automated checks and reviews gate progress, humans handle exceptions and the judgment calls.

Works where the work is well-specified and verifiable — migrations, dependency upgrades, test coverage, mechanical refactors. Fails where intent is ambiguous, which is most of the interesting work.

The economics are real and so is the constraint: throughput is limited by the checks and the review capacity, not by the agent. Adding more generation without strengthening verification produces more plausible work waiting in a longer queue.

How people use it
Planning a large mechanical migration.
"Four hundred files to migrate. That is a factory job: a spec, an agent per file, tests as the gate."
Worked example

The queue that moved to review

Ticket throughput increases sharply after agents are introduced.

The bottleneck simply relocates: now a large queue of unreviewed changes. The next investment has to be verification quality — better tests, automated review, tighter specs — or the apparent speed produces a backlog of unvalidated work.

Sources
  • blogBuilding effective agents ↗Anthropic Engineering · 2024

    Draws the line between fixed workflows and agents that choose their own process, and names the core patterns: chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer.

  • reportState of AI-assisted software development 2025 ↗DORA / Google Cloud · 2025

    Frames AI as an amplifier of the organisation it lands in: the same tooling helps strong delivery cultures and compounds problems in weak ones.

  • blogDemystifying evals for AI agents ↗Anthropic Engineering · 2026

    Gives working definitions for harness, task, trial, grader, transcript and outcome, and argues you must evaluate the model and harness together and judge final environment state, not just the last message.

Related terms