Harness engineering determines whether an AI agent succeeds in the real world: it’s the scaffolding around models—context delivery, tool interfaces, planning artifacts, memory systems, verification loops, sandboxes, and orchestration. This repository gathers canonical essays, practical patterns, and working implementations so engineers can design harnesses that make agents reliable without relying on model upgrades.
What Sets It Apart
- Broad, curated scope — combines foundational essays, design primitives (agent loop, compaction, tool design), concrete reference implementations, and security/sandbox patterns so you don't need to hunt across many repos and blog posts. This saves time when building a production harness and lets you compare approaches side-by-side.
- Practice-first entries — emphasizes engineering artifacts (PLAN.md, AGENTS.md, MCP servers, sandboxes, evals) and template patterns rather than only academic theory, making it actionable for teams shipping agents.
- Cross-vendor perspective — collects vendor docs and open-source projects (OpenAI, Anthropic, Google ADK, LangChain, MCP ecosystems) so you can design portable harnesses and understand trade-offs between provider-specific features (e.g., compaction semantics, extended-thinking APIs).
Who it's for & Tradeoffs
Great fit if you are an engineer or architect building production agent systems who needs a single place to discover: context compaction strategies, tool/MCP design patterns, memory stores, orchestration frameworks, sandboxing options, and CI/eval approaches. The list is practical for internal coding agents, multi-agent orchestration, and safety/governance work.
Look elsewhere if you need a step-by-step installer or a single runnable product—this is a curated bibliography and index of implementations, not a turnkey platform. Also expect some entries to reflect 2026-era practices (MCP, Agent SDKs, microVM sandboxes) and to evolve rapidly; treat the list as a snapshot and a starting point for deeper evaluation.
Where It Fits
Use this as a discovery and design reference when scoping a harness: pick patterns (planner/executor, compaction, permission fabric) and prototype using the referenced implementations (MCP servers, sandbox runtimes, orchestration frameworks) before committing to a cloud-managed product or vendor lock-in.
