A GitHub list of AI agent harnesses now hits 160 entries

RyanAlberts/best-of-Agent-Harnesses ran its weekly rescore on August 16 and came out the other side tracking 160 agent harnesses, up from the 100-plus the project's own description still advertises. The run touched 23 files and swapped +2,229 lines for -2,269, per the GitHub repo.

The count matters less than what sits behind it. The project isn't just a README anymore. It publishes a searchable site with one page per harness, filterable by capability, autonomy, and recovery behavior. It also ships an MCP server with recommend and pick_harness tools, plus an llms.txt and a harnesses.json, so an agent can query the list itself instead of a person reading a table.

That's the shift worth watching. A list built for humans to skim turned into a list built for agents to call. If pick_harness gets used inside real agent workflows, the project stops being documentation and starts being infrastructure other tools depend on at runtime.

The same August 16 rescore hardened the build so an archived upstream project can no longer kill the run. That's a small fix, but it says something about the category: harnesses in this space go stale or get archived often enough that the tracking tool needed a guard against it. A list moving fast enough to need that protection is also a list whose ranking has a short shelf life.

GitHub's page doesn't say how recommend scores harnesses against each other, or what happens when a listed project gets archived after the MCP tools already point at it. Both matter if agents start treating this list as ground truth rather than a snapshot.