Founder of Hermes Labs, an AI reliability engineering lab building tools, research, and autonomous infrastructure for agents and LLM systems.
I work on the operational language layer of AI systems: instructions, tool interfaces, retrieval, memory, prompt injection, runtime policy, evaluation, and the evidence needed to reconstruct what an agent actually did.
Start here: Hermes Labs · GitHub org · LintLang · Research · Contribution ledger
- LintLang: static linter for agent configs, tool descriptions, and system prompts; no LLM call. lintlang.ai · playground
- Little Canary: prompt-injection detection through sacrificial canary-model probes. littlecanary.ai
- Fidelis Memory: agent memory that returns original passages verbatim instead of a paraphrase.
- Hermes Rubric: evidence-first LLM-as-judge scoring; every dimension ties to a quoted file:line, hedged on thin evidence.
- Hermes Blind: recovers the original goal of a long agent session from its first turn, for multi-turn drift recovery.
- zer0dex: local dual-layer memory pattern for agents, pairing a markdown index with local semantic retrieval.
Full catalog with evidence boundaries: hermes-labs.ai/open-source
Hermes Labs has 50+ merged external contributions and ecosystem PRs across AI frameworks, agent infrastructure, developer tooling, documentation, integrations, and public technical systems. The public contribution ledger keeps the categories separate; the underlying record is also available as machine-readable JSON.
- LintLang runs in Character.AI's Larch CI as an operational lint dependency, documented in Larch's own linting docs (scoped to Larch's public repo).
- LintLang is a catalog plugin in MegaLinter, merged upstream as oxsecurity/megalinter#8899 (2026-09-11); a catalog listing, not an install-count claim.
- Independent Gentoo ebuilds for LintLang are maintained across releases in the Haven overlay, packaged without our involvement.
- zer0dex changed an external team's roadmap: AllSource documented how their architecture moved from zer0dex's pattern to their own system (roadmap influence, not a dependency), and Mnemosyne independently implemented a related approach.
- Little Canary's technique was adapted in Preflight with attribution, and named as the closest prior work in ProxyCanary's manuscript (a repository-hosted manuscript citation, not peer review).
- Hermes Rubric is listed in Inspect AI's extensions gallery (accepted upstream docs PR, UKGovernmentBEIS/inspect_ai#5265, merged 2026-09-09).
- A CISA Vulnrichment score correction: CVE-2026-14216 was stored at CVSS 5.3 against a 6.5 vector; a collaborator confirmed the correction and republication in cisagov/vulnrichment#333 (case study). A scoring correction, not a vulnerability discovery, partnership, or endorsement.
Merged fixes in AI frameworks and agent infrastructure, each with a regression test: Microsoft Semantic Kernel #13610 (a chat-history truncation reducer was silently deleting the system prompt), LangChain #35544 (dropped a forced tool_choice that crashed Anthropic extended-thinking requests), DSPy #9978 (an empty eval devset failed with a bare ZeroDivisionError instead of a descriptive one), crewAI #7369 (memory access times were updated even when a read was marked read_only; merged 2026-09-15), and Hindsight #4285 (added task-local retain suspension to the Python client; merged 2026-09-15). A related Mem0 patch fixing a Redis cosine-distance-to-similarity conversion closed unmerged after a maintainer acknowledged the conversion in a broader sweep. It is not a merged contribution.
Many contributions from this account are independently discovered and executed by Hermes Labs' autonomous engineering infrastructure, rather than beginning with me selecting or prompting the specific task. I set objectives, operating constraints, and authorization boundaries; steer or review where needed; and remain the responsible human for work published from this account. The machinery is part of the experiment: can an AI-native engineering institution notice useful work, investigate it, make bounded changes, preserve evidence, and stop appropriately without requiring a human to originate every individual action?
Hermes Labs publishes research and technical notes on AI reliability, epistemic failure, measurement validity, agent-tool semantics, prompt injection, and the limits of model self-report. Current work includes Tool Differentia, Behavioral Canarying for Prompt Injection, The Generative Horizon, Precise Records, Unstable Meanings, A Taxonomy of Epistemic Failure Modes in Large Language Models, and The Asymmetric Burden of Proof.
- Tool Differentia: Relational Static Analysis for AI Agent Tool Descriptions
- Behavioral Canarying for Prompt Injection
- The Generative Horizon
- Precise Records, Unstable Meanings
- A Taxonomy of Epistemic Failure Modes in Large Language Models
- The Asymmetric Burden of Proof
hermes-labs.ai/research · hermes-labs-ai/hermes-publications
Hermes Labs emerged from an exploration of philosophy of language, phenomenology, and hermeneutics applied to AI systems. The practical consequence became an engineering thesis: in agent systems, language is part of the runtime. System prompts, tool descriptions, retrieved passages, memory, summaries, policies, and evaluation criteria do not merely describe a system. They participate in what it notices, chooses, remembers, and does. Hermes Labs treats that layer as an engineering surface: something that can be inspected before deployment, tested under adversarial conditions, controlled at runtime, and verified afterward.