I am the Technical Staff of Myself.
I am an AI agent researcher and engineer building evidence-driven systems around agent memory, completion verification, skill operations, safety, context management, observability, and evaluation.
My work sits between agent research, infrastructure engineering, and hands-on product building. I care about systems whose claims can be traced to public artifacts, repeatable checks, or clearly bounded experiments - not just convincing demos.
Email | X / Twitter | Hugging Face
| Area | What I build | Public evidence |
|---|---|---|
| Agent run evidence and workflow environments | Local-first Run retention, failure diagnosis, controlled instruction comparisons, and resettable workflow Cases with objective validators. | Runtime Evolution Workbench / Workflow Environment Factory / RunCase Interchange / DeepSeek Harness adapter |
| Agent memory and continuity | Local-first memory reconstructed from observable task evidence, with reviewed and redacted promotion instead of automatic transcript dumping. | Agent Memory System |
| Task continuity across lossy boundaries | Source-aware task invariants and bounded recovery across compaction, resume, handoff, and agent changes. | Context Continuity / DeepSeek Harness adapter |
| Intent formation and correction | Local-first current-intent state that keeps user statements, agent inferences, evidence, unknowns, disagreements, and superseded claims distinct. | Intent Loop |
| Completion evidence and replay | Fresh, file-bound verification receipts; fail-closed completion gates; and incident reconstruction from observable events. | BeforeDone |
| Execution fidelity | Contract-bound checks before high-impact actions and before completion claims, while reversible exploration stays quiet. | Execution Fidelity Guard, DeepSeek Harness adapter |
| Skill operations and safety | Methods and tools for turning run evidence into reviewable Skill changes, measuring persistent experience, then designing, testing, operating, and statically scanning reusable agent skills. | Experience to Skill, Governed Skill Evolution, SkillOps, Skill Security Guard |
| Source-backed research workflows | Staged evidence collection, claim discipline, review loops, and publishable long-form research. | Industry Research Framework |
| Project | Why it matters | Explore |
|---|---|---|
| Agent Run Evidence Toolkit | Four independent Apache-2.0 repositories that preserve observable Agent runs, turn failures into reviewable improvement proposals, and convert confirmed workflows into resettable scored Cases. RunCase Interchange is the only shared boundary; the DeepSeek Harness adapter captures the same portable Run format. | Runtime / Factory / Protocol / DeepSeek adapter |
| Agent Memory System | A local-first, cross-agent memory layer that keeps raw evidence local, builds an append-only task ledger, and promotes only reviewed portable memory. Missing or unobservable reasoning is marked explicitly rather than invented. | Repository / Website / Hugging Face |
| Context Continuity | A local-first shared-core plugin that preserves task invariants across compaction, resume, and handoff. It ships as a Codex beta plus a pinned DeepSeek Harness developer-preview adapter, using one state protocol on Windows, macOS, and Linux. | Repository / DeepSeek Harness adapter / v0.2.0-beta.2 / Evidence |
| Intent Loop | A local-first, traceable current-intent layer for long agent tasks. It keeps explicit statements, inferences, evidence, unknowns, disagreements, and superseded claims separate while exposing the same fifteen-tool state core to Codex and DeepSeek Harness. | Repository / v0.2.0-beta.5 / Evidence |
| BeforeDone | A Go CLI and Codex integration that requires current verifier evidence before an agent declares completion. It also supports evidence-only incident replay without claiming access to hidden chain-of-thought. | Repository / Guide / Technical report |
| Execution Fidelity Guard | A lightweight contract and evidence layer for agent actions and completion claims. The Codex release and unofficial DeepSeek Harness alpha share the same seven-field contract while leaving planning, permissions, and execution to their hosts. | Codex repository / v0.2.2 / DeepSeek Harness alpha.2 |
| Industry Research Framework | An agent-agnostic workflow for source-backed industry research: task state, source and claim discipline, depth budgeting, staged drafting, adversarial review, and publication cleanup. | Repository / Read the framework |
| Experience to Skill | A WikiSkill-inspired Codex Skill that keeps selected run evidence, accumulated patterns, and executable guidance separate, then gates one atomic candidate on frozen failure and protection cases before human adoption. | Repository / v1.0.0 |
| Governed Skill Evolution | A prospective three-condition study of no Wiki, flat history, and persistent Wiki across three evolution rounds and three replicates. In this narrow setting, persistence reduced context, latency, and rule length versus flat history, but did not improve mean task quality. | Repository / Paper / v1.0.0 |
| SkillOps | A practical framework for treating modular skills as operated software: design contracts, tests, evaluation, versioning, safety, and lifecycle management. | Repository / Paper |
| Skill Security Guard | A static security scanner for agent skills, with A-F risk ratings, safe archive scanning, CI-tested rules, and machine-readable findings. | Repository / Releases |
- Agent run evidence and workflow environments: Runtime Evolution Workbench v0.2.0 saves observable Codex runs, groups evidence-backed issues, and compares reviewable AGENTS.md or Skill changes; Workflow Environment Factory v0.2.1 turns confirmed repositories and Issue-to-PR flows into resettable, objectively scored Cases. RunCase Interchange v0.1.2 is their only shared protocol, and the DeepSeek Harness v0.1.0 adapter captures compatible runs. Hosted Windows, Linux, and Apple Silicon macOS package lifecycles are public prerelease evidence; authenticated improvement efficacy, physical-Mac use, and a complete Factory Agent run remain unproven.
- Execution Fidelity Guard explores a narrow execution-fidelity layer. Shadow mode reminds and proceeds; balanced mode can deny explicit contract conflicts, return user-owned choices to the Host, and request current contract-bound evidence. The Codex v0.2.2 release and separate unofficial DeepSeek Harness v0.1.0-alpha.2 adapter have source and integration paths in Windows, macOS, and Linux CI. Installed-client UX, online shadow results, and controlled-task efficacy remain pending.
- Context Continuity protects the small set of task facts an agent must not lose when context is compacted, resumed, or handed off. The same local-first core powers the Codex plugin and the DeepSeek Harness adapter on Windows, macOS, and Linux. The v0.2.0-beta.2 prerelease passed its six-job tagged CI matrix; authenticated real-Mac Codex lifecycle evidence and efficacy results remain pending.
- Intent Loop gives Codex and DeepSeek Harness a local, traceable current-intent state without taking over planning or execution. The same fifteen-tool core runs through headless packages on Windows, macOS, and Linux. The v0.2.0-beta.5 prerelease records implementation, privacy, packaging, installed-host, and lifecycle evidence; the frozen paired efficacy study remains
NO RESULT. - Thin Harness, Strong Contracts studies production-oriented agent harnesses as runtime, permission, state, replay, memory, evaluation, audit, and human-review infrastructure. Preprint and artifact.
- Bind important claims to observable evidence.
- Use deterministic checks for deterministic behavior, and model evaluations where behavior is genuinely stochastic.
- Keep sensitive raw evidence local; publish reviewed, redacted, and reproducible artifacts.
- Treat measured results as bounded evidence, not universal proof of efficacy.
- Preserve real failures and user corrections as future regression cases.
Stack: Go | Python | TypeScript | JavaScript | Shell | HTML/CSS