Skip to content
AI|Course

Building agentic systems · Expert

Explore Meta-Harness Architectures for AI Agents

The harness around a model often matters as much as the model. Learn what a harness is, how long-running harnesses are built, what "meta-harness" means, and how to search over harness code with your own data.

7 lessons · 17 hr 30 min · Free

You will learn to

  • ▸Describe a harness as a set of design surfaces you can change and measure
  • ▸Explain the structure of long-running and orchestrator-worker harnesses
  • ▸Distinguish the senses of "meta-harness" and place the research in its lineage
  • ▸Explain the Meta-Harness method, its evidence and its stated limits
  • ▸Build and audit a small harness search with proper data splits and sandboxing

Before you start

  • ·You have built a tool-calling agent
  • ·Comfortable reading research papers and Python code
  • ·Helpful: the Production Agents and Repository Context courses

Syllabus

  1. 01What is a harness?Agent = model + harness, and why the harness matters.1 hr
  2. 02The anatomy of a harnessSeven design surfaces and how to make them optimizable.2 hr
  3. 03Harnesses for long-running workInitializer and worker agents, feature lists and orchestrators.1 hr 30 min
  4. 04What "meta-harness" meansTwo senses of the term, and the research lineage behind it.2 hr
  5. 05Deep dive: the Meta-Harness methodSearch space, proposer, traces versus summaries, and results.2 hr 30 min
  6. 06Build your own harness searchSplits, traces, proposer instructions, selection and audits.2 hr 30 min
  7. 07Capstone projectSearch for a better harness and prove it on held-out data.6 hr