Create an agent that improves other agents (and itself)

[evgenyrp]

The agent should read Hackbot logs and create pull requests with improvements. That's the loop I now run manually locally when testing, but I'm limited to artificial test examples.

It would aggregate the logs and agents' conversation transcripts for the last runs of a specific agent, analyze them for errors and inefficiencies, and propose improvements to the code in a pull request. We can trigger it manually initially. Later on, we can trigger it on a schedule. This would be especially useful to handle edge cases that happen only in production and feature/model drift.

[evgenyrp]

The same agent can also produce a report on how the agents are doing, whether there are any issues, and how useful they are. This is essentially observability + evaluation + fixing, all at once. We could run this automatically, say, weekly for all agents.