AI models get the most attention, but the real moat could be the harness.
Earlier this year, a Meta and Harvard team beat Anthropic at SWE-bench Pro with a cheaper model. Their custom harness wrapped around Sonnet 4.5 scored 52.7%, while Anthropic's own scaffold running the more expensive Opus 4.5 trailed at 52.0%.
With Hermes, the agent's memory of you lives in files on your machine. Your conversations go into a local SQLite database. Honcho runs dialectic reasoning over your past conversations to build a persistent model of how you work.
The agent writes new skills from its own successful runs. Your agent becomes more useful over time as these skill files build up.
Models will be replaced over time, but your harness stays with you.
