Infini-AI-Lab has released Hermit, an open-source framework that trains an adapter around an AI agent while leaving the model and its existing harness unchanged. Its early benchmark results are promising, though mostly self-reported and often based on single runs.
Watch Desk analysis
What happened
Hermit’s first release, v0.1.0, shipped on 25 October, according to MarkTechPost. The Python framework lets developers train external adapters containing instructions, skills, memory, tools or coordination logic, then use them with compatible agents such as Codex, Claude Code and Pi. The project says the approach can adapt agents to new tasks without fine-tuning the model or forking the harness.
Five adapter-evolution methods are available, with two more listed as forthcoming. The team reports that its MetaWorld Model method scored 0.795 on a CL-Bench database-exploration task, against 0.580 for the in-context baseline. Results across the examples use GLM 5.3 Flash; the article notes that most are single runs and self-reported. MarkTechPost’s account also describes a prompt-only result that fell below the baseline, a useful reminder that an adapter is not a performance charm bracelet.
Why it matters
Many agent systems are difficult to modify without disturbing the carefully balanced machinery around them. Hermit’s proposed alternative is to train a separate, inspectable layer and keep successive versions in a Git tree, where they can be reviewed or rolled back. If it works across compatible models and harnesses, developers could experiment with agent behaviour without rebuilding the underlying system each time.
Our read
The interesting idea is the separation of the agent from the thing being trained around it. The early numbers are a reason to investigate, not a verdict: benchmark breadth, repeatability and performance on real work will matter more than one tidy comparison.
What to watch
- Whether independent runs reproduce the reported benchmark gains.
- How well adapters transfer between models and agent harnesses.
- Whether the forthcoming methods arrive with clearer evaluation details.
Discussion spark: Would you trust a separately trained adapter to improve an AI agent without changing its model or harness, or would you want stronger evidence of repeatable gains first?
Sources and evidence
- Infini-AI-Lab Releases Hermit, an Open-Source Harness Adapter Framework for AI Agents – MarkTechPost (29 October 2026, 08:24 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.