AI agents / Field notes

How to Build Reliable AI Agent Harnesses (7 Open-Source Repos)

A narrated HubMesh roundup featuring harness2, qq-bridge, harvey and 4 more, with direct links to every repository.

Good tools are better shared.
Open the code

From the episode to your editor.

What each project does, who it’s for, and what to check first.

Star counts recorded at publication.
google-research

harness2

Runs recursive orchestration to refine prompts and tools across agent execution steps without modifying underlying model weights.

Python21.4k
What to know
Why it matters
Improves agent task performance and iteration quality using prompt and tool refinement instead of parameter fine-tuning.
Best for
Developers optimizing agent workflows without modifying model weights.
Check before adopting
Verify whether reported benchmark improvements replicate across your specific agent setups and target tasks.
Derpyu520

qq-bridge

Routes OneBot v11 QQ chat events and user messages into DeepSeek Harness agent sessions and sends agent replies back.

JavaScript3.2k
What to know
Why it matters
Bridges QQ communication channels with DeepSeek Harness workflows to handle chat-driven agent prompts and approvals.
Best for
Teams connecting QQ channels to DeepSeek Harness agent sessions.
Check before adopting
Explicitly tested against DSH versions 0.2.0-rc.2 and 0.2.1-rc.2; verify compatibility with other releases.
ethanplusai

harvey

Automates sales prospecting and cold email drafting by routing agent operations through local Claude Code or Codex CLI subscriptions.

Python2.7k
What to know
Why it matters
Eliminates per-token API billing by utilizing existing Claude Pro or Max or ChatGPT paid subscription CLIs.
Best for
Sales outreach operators utilizing Claude Code or Codex desktop subscriptions.
Check before adopting
Requires active Claude Code or Codex CLI subscriptions and desktop execution rather than standard API keys.

Records desktop user sessions and leverages GitHub Copilot CLI to convert recorded workflows into reusable agent skills.

TypeScript5.6k
What to know
Why it matters
Allows teams to capture manual computer interactions once and generate repeatable procedural skills for automated agents.
Best for
Developers creating repeatable agent skills from recorded desktop tasks.
Check before adopting
Verify dependencies on GitHub Copilot CLI and supported desktop environments for task recording.

Exposes browser automation tools via MCP to navigate Cloudflare challenges, anti-bot checks, and login walls using nodriver.

Python3.3k
What to know
Why it matters
Enables Model Context Protocol agents to interact with protected websites through Chrome DevTools Protocol automation.
Best for
MCP agents requiring automated browsing past bot-detection mechanisms.
Check before adopting
Check compliance with target site terms and verify ongoing effectiveness against evolving bot defenses.
goodnight000

api-anything

Converts websites into plain HTTP requests by inspecting and learning directly from observed browser network traffic.

TypeScript1.6k
What to know
Why it matters
Simplifies reverse engineering of web endpoints into reproducible HTTP requests without manual browser inspection.
Best for
Developers extracting clean HTTP API calls from website network traffic.
Check before adopting
Verify handling of dynamic authentication headers, payload signing, and shifting undocumented backend endpoints.

Exposes structured enterprise metadata through Model Context Protocol to ground agent queries and prevent schema hallucinations.

Python670
What to know
Why it matters
Constrains agent query generation with canonical enterprise schema metadata delivered over standardized MCP interfaces.
Best for
Enterprise teams connecting MCP agents to verified metadata catalogs.
Check before adopting
Verify compatible enterprise catalog formats and required MCP client configurations before deployment.
Stay curious

Your next idea might be
one repository away.

Explore the channel. Find something worth building with.

Find your next rabbit hole.

ESC

Search across episode titles, topics and featured repositories.