Humes

Notes on agent harnesses

An experimental harness
for coding agents.

Humes is a personal project exploring the instructions around a coding agent, and how to tell whether changing them actually helps.

The current focus is a small evidence loop: observe real work, make the outcome explicit, and test a proposed change against the current version.

The approach

Observe. A local plugin records repository changes during agent sessions. Its hooks read Git metadata, without reading conversations or file contents.

Review. A change needs an outcome, not just a record that it happened. Explicit feedback helps turn an observation into a candidate with a repeatable task and clear evaluation criteria.

Compare. Run the current and proposed versions in separate copies of the same task. Keep the evidence, including inconclusive results. A recommendation still needs a human decision.

Where it stands

Early work, used in my own projects. The current agent adapter runs through Codex CLI. Results apply to the tasks tested; they don’t establish a general improvement.

I’m publishing the approach as it develops. This is a working notebook, with room for things that fail, change, or turn out to be less useful than expected.

Changelog

Field notes, as they happen.

Opening the notebook

A first public description of Humes: what it observes, how changes are compared, and where the work stands. Further notes will follow as the project develops.