honed

honed improves an AI agent's configuration — its context files, skills, hooks, MCP servers, and extensions — from evidence of how the agent actually behaved. It is harness-independent: it reads sessions from Claude Code, Codex, Pi, OpenCode, and Antigravity, and installs configurations back into them.

One learning cycle (ACE — agentic context engineering):

sessions ─▶ reflection ─▶ ratings in the journal ─▶ curation ─▶ candidate ─▶ merge ─▶ new version
  • Sessions are indexed from each harness's transcripts and split into trajectories.
  • Reflection: a reflector agent reviews the configuration against a session — like a code review where the evidence is what the agent did — and returns output in a schema you design, with ratings on the lines it judged.
  • The journal records those ratings per section, together with every change to the configuration and its reason, in a git-tracked workspace.
  • Curation: a curator agent edits the configuration in response to the reflections, one commit per change, with the reason as the commit message.
  • Merging: candidates curated in parallel are combined section by section by a merger agent.
from honed.resources import AgentConfig
from honed.reflection import Reflector
from honed.curation import Curator

agent = AgentConfig.from_installation("codex", "path/to/project", name="graphify", version="1.0")
agent = agent.init_workspace("artifacts/graphify")      # git-tracked, with a journal

reflection = Reflector("codex").handoff(PROMPT, session, agent, SCHEMA)
agent.add_reflection(reflection)                         # ratings → journal

curation = Curator("codex").curate(CURATION_PROMPT, [reflection], agent, scales=[USEFULNESS])
print(agent.diff(curation.candidate()))

Start with installation and the first ACE run. The concepts page defines every concept.