Back to browse
GitHub Repository

Jeju is an experimental local-first agent harness for developers who want to define agent behavior in config, run it with strict boundaries, inspect every effect, and improve it with evaluation evidence.

4 starsGo

Jeju – a local-first agent harness with inspectable runs

by cosmtrek·Jun 5, 2026·1 point·0 comments

AI Analysis

●●SolidBig BrainNiche Gem

Manifest-driven agents with eval feedback loops when most harnesses are prompt-only.

Strengths
  • Append-only trajectory.jsonl provides canonical audit trail for every agent run
  • jeju evolve uses evaluation evidence to iteratively improve agent behavior
  • Config manifest defines permissions, tools, context budget before execution
Weaknesses
  • Windows unsupported, requires DeepSeek API key for the showcase demo
  • Agent harness category increasingly crowded with Cursor, Continue, SmythOS
Category
Target Audience

Developers building bounded AI agents with auditable workflows

Similar To

LangGraph · CrewAI · AutoGen

Similar Projects