This project presents a multi-agent AI coding system designed to tackle complex tasks within a terminal environment. It aims to provide a more strategic and efficient approach to AI-assisted coding by employing an orchestrator agent that manages specialized explorer and coder agents. The system's primary benefit is its ability to achieve competitive performance on benchmarks like Stanford's TerminalBench, demonstrating a novel method for compound intelligence through intelligent context sharing.

How It Works

The system operates with a central orchestrator agent that acts as the "brain," analyzing user tasks and decomposing them into subtasks. It dispatches specialized "explorer" agents for investigation and "coder" agents for implementation, without directly interacting with code itself. A key innovation is the "Context Store," a persistent knowledge layer where subagents deposit discrete, reusable "knowledge artifacts." The orchestrator strategically injects relevant contexts into new tasks, preventing redundant work, reducing context window load, and enabling complex, multi-step solutions. This "smart context sharing" transforms isolated agent actions into a coherent, cumulative problem-solving process.

Quick Start & Requirements

• Installation: uv sync
• Running Evaluations: ./run_terminal_bench_eval.sh
• Prerequisites: None explicitly mentioned beyond standard development environment.
• Resources: Evaluation costs and token usage are detailed for Claude Sonnet-4 and Qwen-3-Coder, with Sonnet-4 using significantly more tokens.

Highlighted Details

• Achieved #12 on Stanford's TerminalBench leaderboard, outperforming Claude Code.
• Utilizes an orchestrator agent that delegates tasks to specialized explorer and coder agents.
• Implements a novel "Context Store" for persistent knowledge sharing and compound intelligence.
• Demonstrates significant token usage differences between models (Sonnet-4 vs. Qwen-3-Coder).

Maintenance & Community

No specific details on maintainers, community channels (like Discord/Slack), or roadmap are provided in the README.

Licensing & Compatibility

The README does not specify a license.

Limitations & Caveats

The system's performance and cost-efficiency are benchmarked, but the README does not detail limitations such as unsupported platforms, specific task types it struggles with, or potential scalability issues beyond the observed token usage differences between models. The project appears to be a personal open-source effort without explicit community governance.