● Systems / 2024 → now
Agent Infrastructure.
How I run agents. The harness on top swaps out: Claude Code, Codex, Cursor, Hermes, Pi, more on the way. The system underneath stays the same. It's one shared brain: a custom skill library, safety-gated MCP forks, durable memory, and review gates, loaded into every tool. Open a different harness and the same setup is already there.
- 1
- shared operating brain
- Swappable
- any harness plugs in
- Safety-gated
- MCP integrations
- Reviewed
- before it merges
§ WiringMany harnesses, one brain
- Claude Code
- Codex
- Cursor
- Hermes
- Pi
- + any harness
The harness is swappable. The brain is the system.
§ LoopNothing merges on an agent’s say-so
- Worktree/parallel-feature-worktreeisolated branch
- Scoped PR/push-prpackaged + opened
- Adversarial read@codexa second agent attacks the diff until no issues remain
- Gatesnpm run verifytypes · lint · tests · build
- Merge/triage-prstriage, re-review, clean up
§01Build stack
What it is built on.
Claude Code / HarnessCodex / HarnessCursor / HarnessHermes / HarnessPi / HarnessCustom skills / Shared brainSafety-gated MCP forks / Shared brainDurable memory / Shared brainWorktree → PR → review / Review loopVerification gates / Review loop
§02Artifacts
The shippable evidence.
Shared skill + MCP library
One library of custom skills and safety-gated MCP forks, symlinked into every harness. The same procedures and gated integrations load in whichever tool I open.
private · shared across harnesses
Review loop
Isolated worktree, scoped pull request, an adversarial second-agent read, then verification gates (types, lint, tests, production build) before anything merges.
private · runtime
§03Adjacent in the grid