Blog

• From Prompt to Context to Harness: Notes on the Agent OS

  2026-05-27

  Coding with AI keeps shifting under us, from prompt engineering to context engineering to harness engineering. A working-through of what a harness actually is, two ways to build an operating system for agents, and a guess at how models might start writing their own skills.

• The Spec Is the Code: A Loop for Agent-Driven Development

  2026-05-23

  When agents write the code, the spec becomes the artifact you actually review. This sketches a spec-exec-eval-bench-improve loop — the format, the agent swarm, and the self-improving harness it would take to make it real — while engaging the spec-driven camp that already named most of the pieces.

• Skills, Tools, MCP, Agents: How I Think About the Pieces

  2026-03-24

  A working taxonomy of the agent stack — tools, MCP, skills, and agents — plus notes on the parts I find more interesting: subagents versus agent teams, and the context- and harness-engineering ideas that seem to decide how far a single setup can scale.

• The Filesystem Is Agent Infrastructure

  2026-03-16

  Three connected ideas about the substrate agents run on: the filesystem as critical infrastructure (and whether a tree is the right shape for it), the benchmark gap for skills and agents that's only now starting to close, and whether we should save prompts instead of generated code.

• Chunk Size Might Beat Model Power

  2026-03-12

  A short experiment idea and the reasoning behind it: a weaker model with better workflow and tighter task boundaries might outperform a stronger one — and a hunch that harness engineering is more about rules than flows.

• Field Notes from Vibe Coding

  2026-03-11

  Learnings from a stretch of vibe coding with Claude Code, skills, and agents — why subagents are usually enough for coding, how complexity breeds regressions, the OpenClaw files-as-memory idea and its failure mode, and a three-role sandbox I'm sketching for an agent that holds personal data.

• Make Regenerating Code Safe, Not Generated Code Good

  2026-01-30

  Vibe coding shines on small tasks but I'm skeptical it scales to large projects. A colleague argues the bottleneck shifts from prompts to context — I'm not even sure that's solvable. My hunch: stop trying to make generated code better and instead make regenerating it safe, with small isolated units and a controlled blast radius.