What this video shows
Dru Knox defines a software factory as a system where agents produce the shipped product while engineers improve the system around them. He orders the target metrics as autonomy, automation, and quality, then describes inner, outer, and meta loops for improving one run, a recurring workflow, and the harness itself.
The practical value comes from separating layers: a control plane records and governs work, an agent-ready stack exposes clear tools and context, and improvement loops turn observed failures into changes. Treat the full software-factory claim as a direction rather than a prerequisite. Teams can apply the loop model to one reviewed workflow while keeping people responsible for acceptance.
Compare the talk with Google's behavioral evaluation guide for coding-agent harnesses, which shows small checks for observable agent behavior.
What you will learn
- An inner loop improves the current agent session, an outer loop automates a repeatable workflow, and a meta loop changes the harness from accumulated evidence.
- A control plane should capture tasks, runs, interventions, results, and ownership before teams try to remove human checkpoints.
- Agent-ready repositories need clear instructions, reliable commands, scoped tools, and tests that explain whether the work is complete.
- Manual takeovers, human review comments, and agent-started pull requests can expose where the workflow still depends on hidden human work.
How to apply this safely
- Choose one bounded workflow where agents already produce acceptable drafts and record every correction, retry, review comment, and failed check.
- Add one behavioral test for the most common failure, such as skipping a required validator or editing outside the allowed scope.
- Automate only the repeatable handoff around that tested task, then keep human review at the merge or release boundary.
- Review aggregate results on a fixed schedule and change the harness only when the traces show a recurring failure or measurable opportunity.
Important limitations
- This is a vendor talk from Tessl, and its final section presents Tessl products. The conceptual loop model can be used elsewhere, but the video does not independently compare platforms.
- The software-factory goal assumes mature tests, observability, permissions, and review. Teams whose agents still need frequent prompting should improve small tasks before expanding autonomy.
Sources to check
- The Anatomy of Harness Engineering Google's September 2026 guide to behavioral tests and regression protection for coding-agent harnesses.
Continue learning on Learnetto
Best coding-agent courses
Learn task shaping, tests, review, and repository context.
Best AI agent evaluation courses
Turn traces and failure modes into regression checks.
AI research delegation lab
Compare sequential and parallel plans before automating more work.