📋 Project Overview
A toy project that reconstructs the supervisor-led multi-agent harness pattern seen in tools like OpenAI Codex CLI and Anthropic Claude Code.
🎯 Problem Definition & Goals
- Problem: Agentic coding tools often hide their orchestration details behind opaque frameworks.
- Goal 1: Recreate a compact supervisor-driven harness architecture.
- Goal 2: Keep schemas, permissions, and event flow explicit.
- Goal 3: Support both CLI usage and programmatic embedding.
⚙️ Key Features & Contributions
- Supervisor-Led Loop: The supervisor dynamically chooses which subagent to call next.
- Explicit Role Boundaries: Planner, worker, and reviewer operate with separate responsibilities and tool scopes.
- YAML Configuration: Provider, model routing, tools, and retry limits are configurable.
- Channel/Event Architecture: Runtime events are emitted through listener and output channels.
- Installable Package: A `tiny-agent` CLI entrypoint is exposed for local-workspace runs.
📈 Results & Learnings
- Inspectable Architecture: Condensed modern coding-agent patterns into a small readable harness.
- Flexible Usage: Supports both interactive terminal runs and embedded library-style use.
- Regression Coverage: Core agent and CLI behavior are covered by pytest tests.
- Key Learning: Reliability depends as much on orchestration design as on model quality.