About The Harness
AI workflows from real work, not demos.
I’m building with Claude across CLI, IDE, and Cowork. This site is where the patterns that survive contact with real work get written down, one pillar at a time.
The audience is practitioners shipping things. Sometimes the thing is a greenfield agent; sometimes it’s a legacy system that needs to keep running while it learns new tricks. Both count. Pillars here aim to land across modern and inherited stacks, regulated and commercial, whatever the constraints.
Each pillar is a durable reference. A short-form video may have brought you here; this page is why the video exists. Start with the founding pillar: What is an agent harness?
What you’ll find here
Six pillars covering the full surface of building with Claude.
- Agent design. Subagents, agent teams, dispatching, role boundaries, coordination patterns.
- Tool design. MCP servers, skills, hooks, tool choice and schemas, tool-use boundaries.
- Prompt and context engineering. Prompt caching, context windows, system prompts, compaction.
- Evaluation and iteration. Eval harnesses, output comparison, regression tests, judging strategies.
- Model choice and cost. Opus / Sonnet / Haiku tradeoffs, caching math, latency vs. quality, cost profiling.
- Real-world integrations. Connecting agents to APIs, databases, services, observability, production guardrails.
Claude Code artifacts (skills, hooks, subagents, MCP) show up as tactical instances inside these pillars, not as separate clusters. Practical tips are welcome when they come from real work; what’s not welcome is listicle bait.
What you won’t find
- Hype takes, AGI speculation, or Anthropic-roadmap guessing
- “Top 10 prompts” slop or thread-farmed tips
- News aggregation or commentary on the news cycle
- Demos disconnected from something that actually ships
New pillar articles when a topic warrants a long-form take; short videos most days.