# AGENTS.md

**Last Updated:** April 9, 2026

## Research Project: Coding Agent Harness Analysis

### Objective
Collect data and feedback on four coding agent harnesses to determine what works best for different model sizes, particularly smaller/local models.

### Harnesses Under Analysis
1. **opencode** - Go-based coding agent
2. **pi** (pi-mono) - Mario Zechner's minimal terminal coding agent
3. **hermes** - Nous Research's agent that grows with you
4. **forgecode** - AI pair programmer with sub-agents

### Data Collection Strategy

#### Performance Benchmarks
- Run terminal-bench and similar benchmarks across all harnesses
- Track relative performance metrics
- Document success rates, speed, and quality of outputs

#### Community Feedback Collection
Feedback organized by harness and model tier:
- `opencode/feedback/localllm/` - Community feedback for local models
- `opencode/feedback/frontier/` - Community feedback for frontier models
- `pi/feedback/localllm/` - Community feedback for local models
- `pi/feedback/frontier/` - Community feedback for frontier models
- `hermes/feedback/localllm/` - Community feedback for local models
- `hermes/feedback/frontier/` - Community feedback for frontier models
- `forgecode/feedback/localllm/` - Community feedback for local models
- `forgecode/feedback/frontier/` - Community feedback for frontier models

### Folder Structure
```
opencode/
  repo/           - opencode-ai/opencode source
  feedback/
    localllm/     - Local model feedback
    frontier/     - Frontier model feedback

pi/
  repo/           - badlogic/pi-mono source
  feedback/
    localllm/     - Local model feedback
    frontier/     - Frontier model feedback

hermes/
  repo/           - NousResearch/hermes-agent source
  feedback/
    localllm/     - Local model feedback
    frontier/     - Frontier model feedback

forgecode/
  repo/           - antinomyhq/forgecode source
  feedback/
    localllm/     - Local model feedback
    frontier/     - Frontier model feedback
```

### Research Focus Areas
- Tool handling and capabilities
- Skills system effectiveness
- Prompt engineering strategies
- Context management
- Error recovery and resilience

### Future Work
Eventually extract best practices and implement improvements specifically optimized for smaller/local models.

### Reference Materials
Files from `../entropy/Research/md/` contain prompt research and strategies:
- `Research.md` - General research methodology
- `Research-prompt.md` - Prompt engineering research and strategies
- `Research-orchestration.md` - Orchestration patterns and strategies

These files contain prompt research and strategies to reference during analysis.