Field noteOct 26, 20261 min read1 source

Infini-AI-Lab open-sources Hermit, a framework for training harness adapters that improve agents without changing model or harness

In one minute

Infini-AI-Lab has released Hermit 0.1.0, an open-source framework for evolving external adapters around an agent harness.

What changed: Hermit adds a distinct open-source approach: training harness adapters to improve agents while leaving the model and harness unchanged.

The project describes these adapters as instructions, skills, tools, memory, or coordination logic, rather than changes to the underlying model or the original harness.

Hermit separates an Evolver, which implements an optimization method, from a Trainer that runs tasks and handles feedback, specification checks, budgets, versioning, and recovery. Version 0.1.0 lists support for GEPA, SkillOpt, MetaWorld Model, Harness2, and Continual Harness. ACE and ReHarness are listed as planned.

The repository includes an OfficeQA example using GLM 5.3 Flash with Codex CLI 0.162.0. It also includes a Lean example in which five Codex agents share a plan and a verified lemma library. The source presents these as examples and supported workflows; it does not establish general performance gains, comparative benchmark results, or that adapter evolution will improve every agent deployment.

Evidence / sources

  1. 1github.com/Infini-AI-Lab/hermit
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicCodexExplore the codex topic hub.
  2. 02topicAgent HarnessExplore the agent-harness topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04source ledgerSource LedgerInspect the public source evidence graph.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract