Scaffolding for agents that can't scaffold themselves

What it does

This is an awesome list—a curated, categorized directory—collecting canonical essays, research papers, and tools about the discipline of “harness engineering.” It organizes writing from OpenAI, Anthropic, Google, Meta, Microsoft, LangChain, and Martin Fowler into sections covering agent loops, context compaction, tool design, MCP integration, memory, permissions, observability, and verification. The goal is to map out the scaffolding that surrounds a model so an agent can run for hours or days without drifting into chaos.

The interesting bit

The list treats the harness as a first-class engineering discipline with a built-in obsolescence plan: every component exists because the current model cannot do it alone, and the best harnesses are designed knowing they will eventually become unnecessary. That is a refreshingly honest framing in a field that usually treats infrastructure as an afterthought to the latest foundation model release.

Key highlights

• Canonical foundations from OpenAI, Anthropic, and Martin Fowler define harness engineering as distinct from prompt engineering and model tuning.
• Design primitives cover the full stack: agent loops, planning, context delivery and compaction, tool design, MCP skills, permissions, memory, orchestration, and observability.
• Production case studies include Meta’s multi-day ML pipeline harness with hibernate-and-wake checkpointing and Microsoft’s Azure SRE Agent, documented as handling over 35,000 incidents autonomously.
• The taxonomy treats evals, verification loops, and human-in-the-loop governance as first-class harness components, not optional extras.

Caveats

• This is a curated reading list and taxonomy, not a framework—expect essays, patterns, and external pointers rather than installable code or original implementations.

Verdict

Bookmark this if you are designing long-running agents, building internal agent platforms, or trying to convince your team that prompt engineering is not enough. Skip it if you want a drop-in SDK or the latest model weights.