As organisations explore software factories and the use of AI to industrialise software delivery, coding agents introduce a fundamental challenge: unbounded intelligence creates an unbounded surface to verify. While that trade-off may be acceptable for everyday development, it becomes problematic when the goal is to produce repeatable outcomes for a specific class of problem.

In this session, you will explore how purpose-built harnesses can bound agent behaviour by separating exploration from adjudication. Using performance profiling and optimisation as the primary example, you will examine a system in which an agent identifies hotspots and proposes fixes while the harness determines whether the improvement is real. Each run either produces a measured gain or leaves the codebase unchanged, with regressions discarded and retried a bounded number of times before the process stops. You will also examine three principles behind this approach: bounded intelligence through exploration and adjudication, invariants and correctness, and trust through inspectable evidence.

What You Will Learn

• Why purpose-built harnesses are useful when repeatable outcomes matter more than open-ended agent behaviour

• How the explore-and-adjudicate model separates agent actions from outcome verification

• How invariants, correctness checks, and inspectable evidence contribute to trust in AI-assisted workflows

Who Should Attend

Software Engineers, Technical Leads, Staff and Principal Engineers, Software Architects, Platform Engineers, Engineering Managers, and practitioners exploring AI-assisted software delivery and software factory approaches.