53 Free Agent Skills That Check Your AI's Work

Christoffer Tyko describes the application harness as the part you do not control, such as the memory, chat session, hooks and rules that live inside Claude Code. The user harness is the side you do control, and he says the key to making it work is skills, which he treats as instructions, the way you would brief a capable colleague (4:30).

He explains that AI systems are probabilistic, so you will never get exactly the same outcome every time. Agents increasingly run commands through Python or bash scripts, because a script always gets the same outcome (5:58). Later he adds that you cannot make the AI deterministic, but you can give it a deterministic tool set around it (36:24).

Christoffer Tyko says language models are trained mostly on reward and punishment, and adversarial review uses that by rewarding the model for finding lies, discrepancies and bugs. The reviewer compares what the team built with the original specification and tries to tear the design down (13:20). He says this only works when an intake step has first produced a clear start and a clear goal to compare against (17:34).

He says that without hooks there is no guaranteed way of executing the Python scripts, and you do not want the model to decide when a hook is triggered. Hooks sit between the model and the operating system, so commands pass through them first, and because they run outside the agent it cannot circumvent them (22:36, 42:38).

He saves the whole run history as files in a folder in the codebase, including intake files, artifacts, designs and revisions. Because the memory lives in the project and not in the application, a crashed tool can resume from a checkpoint and the work can continue in a different coding agent (25:50).

After the initial conversation and the intake, Christoffer Tyko says the responsibility passes to the Admiral. A failed step is first revised within the team, then by the Admiral, and each step gets three tries. If the same step fails three times, the person becomes the decision maker (45:27, 46:17).