Harness engineering for AI agents.
I build the scaffolding that makes agents reliable — the prompts, skills, memory, and feedback loops around the model, not the model itself.
Repos · Featured work · What I think about
Most of an agent's behavior isn't the model — it's everything wrapped around it: the system prompt, the tools it's given, the skills it can load, what it remembers, and how it recovers when it's wrong. That layer is usually invisible and untested. I work on making it inspectable, reusable, and self-correcting.
Based in Seattle. Currently building always-on autonomous agents, the harnesses that keep them honest — and the tools that audit what those agents believe.
memory-doctor — Audits the quality of your AI agent's memory: what's stale, contradictory, never-loaded, or never-used. Existing tools clean up disk and tokens; memory-doctor audits truth — dead references, entries past the silent load cliff, relative dates that mean nothing at recall time, index orphans that make memories invisible. Local-first, read-only, zero dependencies. First run on my own 13-agent setup found 17 problems, including 20+ paths my agents believed in that no longer exist. Every detector ships with golden fixtures and gets precision-tested on a live fleet before release.
harness-loop — A self-improving loop for agent harnesses. Every prompt, config, memory write, and hook ultimately compiles into one artifact: the HTTP request sent to the model. harness-loop taps that request, checks the compiled prompt against rules you declare, and runs a detect → notify → repair → re-verify cycle so "my memory system works" becomes a fact, not a guess. Zero dependencies, works with any Anthropic- or OpenAI-compatible agent. Born from running three always-on Claude Code agents through this exact loop.
Together they cover both halves of agent reliability: harness-loop audits what the agent does (mechanisms), memory-doctor audits what it believes (memory).
agent-workspace — A local-first handoff protocol so work survives the agent that started it. When a coding agent runs out of context, gets swapped for another model, or hands back to you, the thread usually snaps — the next agent re-derives state from a stale chat summary and quietly redoes or breaks things. agent-workspace makes a .agent-workspace/ directory (state, tasks, decisions, a runlog, and verifiable claims) the source of truth instead, and ships agentws doctor, which scores handoff-readiness 0–100 so "the next agent can pick this up" is measured, not hoped. Python stdlib only, zero dependencies. Dogfooded across the repos on this list — while scoring them the doctor flagged its own false-positive, which became the very next commit.
If harness-loop and memory-doctor are the two halves of one agent's reliability, agent-workspace is the third axis: continuity across agents and context resets.
skills-lab — A library of reusable Agent Skills for Claude Code, plus the authoring principles behind them. Covers research, code/design review, incident response, and an oncall-agent that composes a set of specialist skills into one assistant. As much about how to write skills an agent will actually invoke correctly as the skills themselves.
.dotfiles — Tools and settings for a fast agent-driven dev environment.
- The compiled prompt is the source of truth. You can't verify a harness by reading its config — you verify it by inspecting the request that config produces.
- Memory you can't audit will quietly go wrong. Agents accumulate beliefs; without staleness, contradiction, and usage checks, yesterday's fact becomes today's confident mistake.
- Structural over remembered. Sensing and verification should be things the agent can't forget to do, not instructions it might skip.
- Work should outlive the agent that started it. Context windows end, models get swapped, sessions reset — continuity has to live in files the next agent reads, not in a chat history it never saw.
- Skills are progressive disclosure. A good skill's
descriptionearns its place in context; its body is written for a model to act on, not a human to admire. - Judgment belongs to the agent; approval belongs to the human. Automate the loop, gate anything durable or outgoing.
- GitHub: @jingchaodev
- X: @cifra1902