HALO (Hierarchical Agent Loop Optimizer) provides a methodology and Python package for recursively self-improving AI agent harnesses using Recursive Language Models (RLMs). It targets developers building and deploying AI agents, offering automated optimization by identifying and rectifying systemic failure modes, thereby enhancing agent performance and reliability in production environments.

How It Works

The core approach centers on a recursive loop: collect execution traces from agent harnesses using OpenTelemetry-compatible tracing. These traces are fed into the specialized HALO-RLM engine, which decomposes them to identify common failure modes and systemic issues. A coding agent then uses the RLM's findings to generate and apply changes to the harness. This RLM specialization is advantageous as it avoids overfitting to individual traces, enabling generalization across harness-level problems, unlike general-purpose models.

Quick Start & Requirements

Install via pip: pip install halo-engine. Requires an OPENAI_API_KEY environment variable. CLI usage is demonstrated via halo --help and example commands. Demo projects are available for exploration. Development setup involves git clone, uv, and go-task.

Highlighted Details

Achieved significant performance gains on the AppWorld benchmark: Gemini 3 Flash dev SGC improved by +15.8 points (36.8% to 52.6%), and test_normal SGC by +10.7 points (37.5% to 48.2%). Sonnet 4.6 dev SGC improved by +15.8 points (73.7% to 89.5%), and test_normal SGC by +10.7 points (62.5% to 73.2%). Identified specific harness failures like hallucinated tool calls, redundant arguments, refusal loops, and semantic correctness issues, with findings independently verified.

Maintenance & Community

Contributions are welcomed. No specific community channels (e.g., Discord, Slack) or details on maintainers/sponsorships are provided in the documentation.

Licensing & Compatibility

Released under the MIT License, which is generally permissive for commercial use and integration without copyleft restrictions.

Limitations & Caveats

No specific limitations, alpha status, or known issues are detailed in the provided documentation.