<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> This repository provides the reference implementation for Continual Harness, a reset-free framework designed to automate the refinement of agentic harnesses through online in-context learning. It enables foundation agents to self-improve by dynamically rewriting their prompts, sub-agents, skills, and memory mid-episode. The project also includes the PokéAgent benchmark infrastructure, which has enabled AI agents to complete multiple Pokémon RPGs, serving as a robust platform for evaluating agentic capabilities in complex, long-horizon tasks.

How It Works

Continual Harness operates by allowing a "LLM Refiner" to perform CRUD edits on an agent's harness components—system prompt (p), sub-agents (G), skills (K), and memory (M)—in place during a single, continuous episode. This dynamic refinement occurs via edits on a trajectory window every --optimization-window-length steps, identifying failure signatures and rewriting harness elements without requiring environment resets. This approach extends to joint training of open-source model weights using an online DAgger and process-reward-model pipeline, offering a novel method for continuous agent adaptation and skill acquisition.

Quick Start & Requirements

Installation involves cloning the repository, setting up a Python environment using uv sync (recommended) or pip install -r requirements.txt, and installing the mGBA system library for Emerald. Key prerequisites include Python 3.10-3.11, legally obtained Pokémon Emerald and Red ROMs, and API keys for supported VLM backends (OpenAI, Gemini, Anthropic, etc.). The project offers a web UI for real-time streaming at http://localhost:8000/stream.

Highlighted Details

• Continual Harness Scaffold: Implements reset-free online adaptation, allowing agents to dynamically evolve their harness (prompt, sub-agents, skills, memory) mid-episode via the evolve_harness tool.
• PokéAgent Benchmark: Features a comprehensive benchmark harness for AI agents completing Pokémon RPGs, notably achieving the first AI completion of multiple such games.
• Multi-VLM Support: Integrates with a wide array of Vision-Language Model backends, including OpenAI, Google Gemini, Anthropic, and OpenRouter.
• External CLI Agent Integration: Supports external agents like Claude Code, Gemini CLI, Codex, and Hermes through a Message Communication Protocol (MCP) proxy.

Maintenance & Community

No specific details regarding maintenance, notable contributors, sponsorships, or community channels (e.g., Discord, Slack) were found in the provided README text.

Licensing & Compatibility

The specific license for this repository was not detailed in the provided README text.

Limitations & Caveats

Performance gains from Continual Harness are capability-dependent, showing significant benefits on high-capability models like Gemini 3 Pro but potentially underperforming on less capable models. Users must legally obtain the required Pokémon ROMs, as they are not included.