agent harness

A local-first agent harness for running 2–9B models on laptops with Ollama. The project focuses on finishing real tasks, includes an offline mode, benchmarks, and a large public test suite.

Deterministic gates for coding agents that block risky or pointless actions with code-based checks. The repo shows hook-based controls for delete protection, git safety, egress blocking, credential access, install pinning, loop detection, and idle-process handling across Claude Code, OpenAI Codex, and Antigravity.

A verification browser for AI coding agents built on WebKitGTK, aimed at fast visual checks, page assertions, and human handoff. The repository shows install steps, CLI usage, MCP support, and a concrete agent workflow.

VAEN is a CLI for packaging AI coding-agent setups into portable .agent archives and importing them across repos or teams without copying instructions, skills, MCP config, or secrets.

OpenBot is a local-first harness for running AI agents. It includes a small event API, local file storage, a built-in orchestrator agent, a deterministic state agent, and support for custom agents and plugins loaded from disk.

Research code for improving agent harnesses through regularized recursive self-improvement, with evaluation, selection, pruning, and domain-specific adapters.

Keel is a Rust CLI harness for AI-assisted coding workflows. It adds auditable stopping conditions, durable memory, projections into agent instruction files, and reviewable gates with exported evidence.

agentpack is a tool for staging reproducible AI coding environments across agent harnesses like Claude Code, Codex, Cursor, and OpenCode. It manages skills, hooks, MCPs, and other agent configuration from a single manifest without polluting the project or user config.

An open-source multiplayer agent harness for work, designed for Slack and web-based collaboration with scoped memory, permissions, sandboxes, and admin controls.

Retriever AI’s blog post explains a browser-agent architecture built around DeepSeek Flash, text-only DOM handling, and executable code as the planning layer. It includes concrete examples of the harness, the model’s role, and the cost argument behind the approach.

Open-source AI agent harness tool that lets an agent wipe its session context and replay the first message, designed for loop-based agent workflows.

A self-hostable Slack-first agent harness that turns plain-English requests into specs, code changes, adversarial review, and pull requests, with local Docker and Fly deployment paths.

A detailed article about the internal harness Nimbalyst built around Claude Code and Codex, covering context, provenance, capability, workflow, restraint, verification, visual interface, and coordination.