I integrate coding agents into your repositories. Harness engineering defines their instructions, permissions and the checks your team can rerun.
The method
The pilot tests a new way of working on a real task. Agents carry out the agreed steps, and the team designs the checks and keeps control of decisions.
We start with a task in your repository. We agree on what the agent may change and who approves the result.
We set up instructions, tools and access. We deliberately go outside the agreed scope to check that the safeguard stops the change.
Your team reruns the checks and decides what to accept, adjust or stop.
If a check fails, we adjust the harness, then rerun the scenario.
In an anonymised delivery case, a local replay with 264 synthetic IDs showed the SQL query going from 530 joins to 3. An expiry test checks that the access rule holds as time advances. Read the delivery case (in French)
What your team can do on its own
Update instructions and access permissions, rerun the verification scenarios and decide what to do when a check fails.
I build and evaluate harnesses: context, instructions, tools, permissions, tests and execution traces. I work directly in the repositories.
Deliverable
A maintained agent environment, evaluation cases and a handover prepared with the person taking ownership internally. Before I leave, we review the handover together: they need to be able to identify the next action and know who to contact if they get blocked.
I establish a baseline, run a bounded pilot and review the results with the team. Training and mentoring support adoption.
Deliverable
A pilot assessment, adoption playbook and a decision to extend, revise or stop.
I define permissions, accountability, exceptions and review practices with engineering, security and leadership.
Deliverable
Usable policies, decision records and checks that fit daily work.
Context engineering starts with a task in a repository. We choose the sources and instructions the agent needs, where the team maintains them and how to catch a stale reference.
I built ctxharness to check facts in files that agents read, such as paths and scripts. In your team, we decide which instructions to check, when to run the checks and who fixes a failure.
Explore ctxharnessIt is the environment the agent works in. It brings together the context, authorised tools, permissions and checks that let your team verify its output.
The tested configuration, verification scenarios and handover instructions. We review them with the person taking ownership.