Lead AI Product Engineer · Agent Governance · Verifiable Autonomy
Bicameral · MythologIQ Labs · Microsoft Agent Governance Toolkit · AgenTrust
Governance by prompt is not real governance. I build agent systems where authority, provenance, validation, and release decisions are enforced outside the model's own reasoning.
I build trustworthy autonomous-agent systems across four connected areas:
- Production agent systems: governed local-first runtime, MCP, evidence ingestion, authority boundaries, deterministic validation, and release evidence at Bicameral.
- Open-source governance: listed maintainer on Microsoft's Agent Governance Toolkit, with work spanning policy enforcement, delegated-context governance, contributor trust, MCP verification, and security controls.
- Verifiable agent state: contributor to AgenTrust Agent Manifest, including governed persistent-memory checkpointing and fail-closed state-evolution semantics.
- Independent systems: local-first agent products, governed memory, agent-driven software development, privacy, and reusable AI skill infrastructure.
| Area | Contribution | Evidence |
|---|---|---|
| Microsoft Agent Governance Toolkit | Governed accumulated context across delegated multi-agent workflows; contributor trust and MCP verification work. | Maintainers · PR #2800 · PR #2852 |
| AgenTrust · Agent Manifest | Designed the v0.2 incremental memory checkpoint/delta binding protocol for governed persistent state. | Issue #174 · PR #190 |
| Agent Memory | Public reference architecture for governed memory, lifecycle, forgetting, provenance, conformance, and runtime evidence. | Repository · ADR-021 |
| Bicameral | Governed acquisition, redaction, provenance, decision evidence, and fail-closed validation in a production agent stack. | PR #269 · PR #287 |
| skillz | Passive, provenance-aware skill knowledge resource for agents, with first-party skills and exact-version third-party review evidence. | Repository |
- Uncertainty may propose. Authority constrains. Durable consequences require bounded permission.
- Provenance over phrasing. Where an instruction, memory, artifact, or decision came from matters more than how confidently it is worded.
- Negative paths are part of the contract. Replay, stale state, residual data, partial failure, and unauthorized mutation must have defined behavior.
- Evidence should outlive the agent. Important decisions need inspectable artifacts, not a claim that the model considered them.
Languages: Python · Rust · TypeScript · Go
Agent systems: MCP · orchestration · local inference · RAG · governed memory
Security / trust: policy engines · capability mediation · provenance · redaction · tamper-evident evidence
Platform: GitHub Actions · Docker · PostgreSQL · SQLite · FastAPI · Tauri
MythologIQ Labs, LLC · governed agents · verifiable trails · accountable autonomy