Harness engineering
Two harnesses, one way of working
An engineering harness keeps AI-written software sound as the pace goes up: an agreed way of working, automated checks, and a record of what changed and why. We run two. Both apply the same checks. The difference is where the engineer sits.
What every harness enforces
Gates fail closed
A change that cannot produce evidence it is sound does not merge. Anything the checks cannot settle goes to a human, never through.
Evidence travels with the change
Every merge carries a record of what was checked, by what, and with what result.
Standards live in configuration
The testing, small diffs, and traceability your team agrees on are written into configuration, so they apply to every change.
Spend stays visible
Each change can record which model did the work and what it cost, and routine steps can run on smaller, cheaper models. You can turn on both in either harness.
Human in the loop
LightFactory
A human reviews the research and approves the plan before any code is written. The agent does the reading and the drafting, and the two decisions that set direction stay with a human. Most of our engagements run this way.
- Research
- Human reviews
- Plan
- Human approves
- Implement
We use it when
You are starting from a prototype or an early idea, the codebase is unfamiliar, or the cost of building the wrong thing is higher than the cost of a slower start. LightFactory turns a working prototype into production software without losing what the prototype taught you.
What you get back
A reviewed research note and an approved plan for every change, so the reasoning behind each decision stays on record after the people who made it move on.
Human in the control tower
HydraFlow
A staged pipeline moves each issue through gates, and a human oversees every gate from above.
- Triage
- Plan
- Implement
- Review
- Merge
Coverage and CI gates apply at every step. Anything ambiguous escalates to a live queue for a human.
We use it when
The work is well understood and repetitive at volume: a migration, a platform port, a backlog of similar changes across many teams.
What you get back
More changes shipped than hand review allows, with the same record of what was checked along the way.
Result · Fortune 10 Ads
15 min
to production for an experiment, down from two weeks.
Proof · Fortune 10 Ads Platform
Production releases in 15 minutes, down from two weeks.
In an advertising unit that generates billions of dollars a year, getting an experiment into production took two weeks. We built an experimentation platform that gets an experiment into production in 15 minutes, safely.
Questions
What buyers ask us
What is harness engineering?
Harness engineering is building the rules, checks, and records that let AI agents write production software safely. Our view: the model matters less than the harness around it. Models change every few months. A good harness keeps your standards in force through every one of those changes, so your team gets faster without the codebase getting worse.
Is this lock-in?
No. The harness lives in your repositories, your pipeline, and your cloud, so your team can see and change every rule and check. We build it so your team can keep running it after the engagement ends.
How do we choose between them?
With us. The two differ in where the human sits, and that changes speed, supervision cost, and risk. We help you weigh those tradeoffs against your codebase, your team, and the work in front of you, then build the delivery process around the choice. Some clients land on one. Others run LightFactory for new work and HydraFlow for a large migration.
Is a harness only pipelines and tooling?
No. A harness has two parts: the automated checks in your pipeline, and the standards and decisions your team agrees on before anyone writes code. The agreements matter as much to quality as the checks do.
What if our codebase is a mess?
Then that is where we start. An agent copies the patterns it finds, so inconsistent code gets reproduced faster. We help you choose the conventions that matter, bring the parts of the codebase that change most into line with them, and build the harness as that work lands. Most teams start with one small project in the area that changes most, prove the approach there, then widen it.
Who you will work with
Travis Frisinger Head of Agentic AI
Alex Haldeman Principal Engineer
Stephen Walker Senior Crafter
Related
What would a harness look like on your codebase?
Tell us how your team works today, and we’ll walk through where a harness would help first.