Back to blog

Kasetto

How a versioned config makes an AI agent harness repeatable across projects and coding agents.

The harness is part of the engineering system

A coding agent starts with more than a request. It reads project instructions, discovers skills, connects to tools, and may call specialized agents. If each engineer assembles those pieces by hand, two people can ask the same agent to work in the same repository and give it different context.

I use harness to mean the managed setup around an agent: instructions, skills, tool connections, commands, agent roles, and plugins. The goal is not to make a model smarter. It is to make the capabilities and operating context predictable enough for a team to rely on them.

Keep the setup in one project declaration

Kasetto puts that declaration in kasetto.yaml. The file can name the target agents, the scope of an install, and the sources for skills, MCP servers, commands, subagents, instructions, and plugins. The repository then carries a reviewable account of what its agents can load.

A project-level config can target more than one coding agent from the same source. That is useful when a team uses different hosts but wants the same project guidance. The native files still differ by host, so a manager needs to translate the declaration into each host's expected layout.

scope: project
agent:
  - codex
  - claude-code

skills:
  - source: ./agent-skills
    skills:
      - code-review
      - release-check

Make sync a reviewable operation

Kasetto's sync command reads the config and lock file, resolves the declared sources, and updates the installed resources. It uses hashes to tell whether tracked skills have changed. MCP entries are merged into native settings files, and commands or subagents can be transformed into the format a host expects.

The dry-run option previews planned changes before Kasetto writes them. The lock file records resolved sources, so a normal sync can reproduce the selected versions. Updating a source is a separate, intentional step. That gives reviewers a clear diff when a team changes an instruction pack or moves to a new revision.

Keep setup separate from task execution

Kasetto prepares the files an agent host reads. It does not decide which workflow should handle a request, run a skill's instructions, or perform the work itself. Once the host starts an agent, that host interprets the installed setup and carries out the task.

That boundary matters. A setup manager should make configuration consistent. A router should choose a workflow. An orchestration process should coordinate execution. Keeping those responsibilities separate makes it easier to see where a change belongs and who owns it.

Treat the harness as an owned product

For an engineering team, a shared config is a small but useful control point. Platform owners can review new sources, choose project or global scope, and make sync steps part of onboarding. Engineers can see which resources a project expects instead of copying setup instructions from an old chat or wiki page.

The config still needs an owner. A large catalog of stale skills is not a standard. I would start with a short set of resources tied to real work, review updates like other dependency changes, and check that sync writes the expected native files. That turns agent setup into an operational capability the team can maintain.

References

Kasetto documentation

How Kasetto syncs resources