I build the operating systems around coding agents: routing, durable project state, verification, deployment coordination, and the boundary between agent confidence and actual evidence.
My flagship project is Citadel, the open-source operating layer for Claude Code and OpenAI Codex.
one command
↓
route the work
↓
verify what happened
↓
write state and evidence into the repository
↓
resume in a fresh session or coordinate isolated agent fleets
| Result | What it demonstrates | Inspect it |
|---|---|---|
| Crossed 780 GitHub stars during this surge | Developers are actively discovering the operating-layer problem | Citadel |
| 107 unique cloners in one day | Attention reached repository acquisition, not only stars | Activation methodology |
| 15 of 15 PRs merged and deployed | One steward protected main while other agents kept producing | Public proof repository |
| 14 branch updates, 59 CI waits, 0 repairs | The landing lane handled real merge and CI pressure without inventing success | Deploy steward proof |
| 30 of 30 hosted journeys | Claude Code and Codex complete the same operating journey across Windows, Linux, and macOS | Golden path |
| Citadel v1.1.0 | Reproducible archive, manifest, checksum, update plan, and rollback path | Release |
These are bounded claims. Deterministic fixtures are not human adoption, clone operations are not retained users, and missing evidence remains unknown.
Coding agents are already capable of meaningful work. The unsolved problem is operating them across real projects:
- choosing the right amount of machinery for each task
- surviving context resets and fresh sessions
- coordinating multiple agents without sharing one branch
- protecting files and approval boundaries
- distinguishing completed work from unsupported confidence
- leaving enough evidence for the next human or agent to continue safely
Citadel makes those operating contracts repository-native and inspectable.
Persistent project memory, /do routing, safety hooks, cost telemetry, verification receipts, multi-session campaigns, and parallel fleets for Claude Code and Codex.
Standalone Claude Code skills forged from real engineering problems.
A public stress test of serialized post-PR coordination through protected main.
A verifiable AI agent that judges work submissions on EigenCompute.
I am currently improving Citadel around four questions:
- Can a stranger install it and reach one verified success in under ten minutes?
- Which workflows create enough value that developers return after seven days?
- How should coding-agent systems represent missing or contradictory evidence?
- How can multiple agents keep producing while one lane safely protects main?
If you are operating Claude Code or Codex on a real repository, I want to hear where the workflow breaks.
- Try Citadel
- Share a workflow or failure in GitHub Discussions
- Follow development on X
- Connect on LinkedIn
I am especially interested in agent infrastructure, developer tooling, verification systems, orchestration, and teams trying to move from impressive demos to dependable operating practice.