$ find your agent harness

MIT-licensed Python library (103k stars) that lets any LLM drive a real Chrome browser via DOM+vision instead of brittle selectors, now flanked by a desktop app and a usage-based cloud with stealth browsers, proxies, and a 24/7 'Browser Use Box' agent.

OpenAI's open-source (Apache-2.0) Rust-rewritten terminal coding agent, distinguished by kernel-level OS sandboxing (Apple Seatbelt on macOS, Landlock/seccomp on Linux), a 272K-1M token context with GPT-5.x Codex models, MCP support, and cloud handoff to Codex web tasks.

Formerly OpenDevin, the leading open-source autonomous software-engineering platform: Docker-sandboxed agents that write code, run commands, browse, and open PRs, with a production agent SDK hitting 72% on SWE-bench Verified (Claude Sonnet 4.5) and its own broader OpenHands Index benchmark launched Jan 2026.

Mario Zechner's radically minimal MIT/TypeScript harness (~68k stars): four tools (read/write/edit/bash), a ~300-word system prompt, deliberately no built-in MCP/subagents/permission-popups, with everything else added via TypeScript extensions, skills, and npm/git-shareable packages across interactive, print/JSON, RPC, and SDK modes.

MIT-licensed Python framework (~55k stars) for role-based multi-agent 'Crews' plus event-driven 'Flows', with unified memory, hierarchical delegation, and a commercial AMP cloud platform (free tier, Pro from ~$25/mo, enterprise) that CrewAI says powered ~2B agentic executions in the year to January 2026.

Block's Rust-built, Apache-2.0 general engineering agent — now stewarded by the Linux Foundation's Agentic AI Foundation — distinguished by the deepest MCP integration in the ecosystem (70+ extensions, can also act as an MCP server) and shareable YAML 'Recipes' that scaled agent workflows to 60% of Block.

Low-level graph-based agent orchestration runtime with automatic checkpointing, durable execution, human-in-the-loop interrupts, and streaming (v1.2 in May 2026 added typed streaming, per-node timeouts, and distributed runtime), used in production by Klarna, LinkedIn, Uber, and Replit and serving as the substrate under Deep Agents.

Hugging Face's minimalist (~1,000-line core, Apache 2.0) 'agents that think in code' harness where the model writes executable Python as its action format instead of JSON tool calls, model-agnostic via transformers/Ollama/LiteLLM, with optional sandboxed execution (E2B, Docker, WebAssembly), managed sub-agents, and Hub sharing of tools and agents.

The original open deep-research harness (Apache-2.0, 28k stars): a planner/executor multi-agent pipeline that fans out parallel web and local-document searches into cited reports, with recursive tree-like 'Deep Research', MCP data-source integration, and any-LLM support.

OpenAI's agent harness whose April 15, 2026 overhaul added a 'model-native harness' with Codex-like filesystem/shell tools, built-in compaction and memory, snapshot/rehydrate durable execution, handoffs, guardrails, tracing, and native sandbox execution via bring-your-own or built-in providers (E2B, Modal, Daytona, Cloudflare, Vercel, Blaxel, Runloop); non-OpenAI models supported via LiteLLM and Chat Completions-compatible endpoints.

The leading TypeScript agent framework (Apache 2.0 core since July 2025; some enterprise features commercial) offering agents, type-safe workflows with suspend/resume human-in-the-loop, memory, a unified router to 3,300+ models across ~94 providers, guardrails/scorers/tracing, Mastra Studio for interactive testing, and file-based agent definitions added July 2026.

LangChain's self-described 'batteries-included agent harness' (GA March 2026, ~26k stars) for long-running tasks, layering planning via write_todos, a virtual filesystem for context offloading, SKILL.md-compatible skills with progressive disclosure, subagent spawning, and LangGraph-store-backed long-term memory on top of the LangGraph runtime.

Alibaba's Apache-2.0 terminal agent (originally a Gemini CLI fork) optimized for Qwen3-Coder models, distinguished by broad provider flexibility (OpenAI/Anthropic/Gemini/Qwen APIs plus Ollama/vLLM local models), skills/subagents primitives, headless mode, and TypeScript/Python/Java SDKs.

Apache-2.0 platform for stateful agents descended from MemGPT, distinguished by tiered self-editing memory persisted in a database (core memory in-context, archival out-of-context) plus 2026 research on git-based 'Context Repositories' and memory-native RL; in 2026 the company is consolidating around Letta Code, an open, model-agnostic memory-first coding agent harness (LettaBot archived May 2026 and folded in as Channels).

MIT-licensed, type-safe Python agent framework from the Pydantic team with a FastAPI-like developer experience: schema-validated structured outputs, first-class MCP support, built-in OpenTelemetry observability (pairs with Logfire), and durable execution via native Temporal/DBOS integrations that offload model requests, tool calls, and MCP communication to fault-tolerant activities.

MIT-licensed open agent harness from HKU's Data Intelligence Lab (agent loop, tool registry, permission system, hooks, skills, plugins, MCP, React TUI, automatic context compression) built for clarity and Claude-style workflow compatibility, shipping with 'Ohmo', a built-in personal agent that runs on existing Claude Code/Codex subscriptions and works via Feishu/Slack/Telegram/Discord, autonomously branching, coding, testing, and opening PRs.

The AutoGen + Semantic Kernel convergence (1.0 GA April 2026) for .NET and Python, whose BUILD 2026 'Agent Harness' made production patterns first-class — shell/filesystem access, human-in-the-loop approvals, context compaction, todo tracking, plan/execute modes, AgentSkillsProvider skill discovery, and BackgroundAgentsProvider subagents — with hosted execution via Microsoft Foundry.

Anthropic's production agent library exposing the same harness that powers Claude Code (tool loop, file/bash/web tools, subagents, hooks, skills, MCP, permission checkpoints) in Python and TypeScript; the MIT-licensed SDK wraps the bundled Claude Code runtime and, as of 2026, underpins Anthropic Managed Agents with separate metered billing from June 15, 2026.

A research-grade 'harness foundry' (MIT, v0.1.0 beta, arXiv paper 2606.14249) that composes agent harnesses from typed reusable processors across seven categories (context, control, evaluation, memory, multi-model, observability, tools) and evolves them via the AEGIS trace-driven engine, reporting an average +14.5% gain across ALFWorld, GAIA, WebShop, tau^3-Bench, and SWE-bench Verified without touching model weights.