Skills capture the know-how of a service or organization. They donโt define the whole agent: its model, tool access, execution logic or expected results.
Versioning skills alone leaves critical agent behavior outside of an engineer-approved review process. This in turn leads to inconsistent agent usage & unpredictable token costs.
Put every agent definition, including your coding agents, in Git so that your team can:
Review instructions, models and tool access changes as PRs.
Run evals to ensure skills are indeed being used as expected and trace regressions to specific changes.
Compose agents into multi-agent workflows.
Combine LLM judgment with deterministic code and control flow.
Your coding agents can create, test and improve those definitions too.
Every engineering team needs agent definition to be explicit, versioned and reviewable.
