Art director and creative technologist in Los Angeles. I direct AI video and build the agent tooling behind it.

One man walks through a door on a Los Angeles sidewalk and keeps walking, through eighteen worlds. A 78-second reel built around the hardest problem in AI video: keeping one character the same from shot to shot.

Watch: isaacsight.com/many-worlds.html · Bibles, prompts and edit code: isaacsight/many-worlds

More work: reels · one-page overview · how I direct AI: the studio rules

Provenance engineer. Editor at kernel.chat. Building open-source audit-grade infrastructure for AI agents in regulated industries.

I work at the seam between AI agents and the deterministic engines they have to call to stay legal. The architectural rule I commit to: the AI never produces the source-of-truth number — engines do, humans approve, every step is content-addressed and hash-chain logged. That's the work.

POSIX for AI agents. Ten OS-level primitives — signed capability tokens (acap), namespaces, per-agent quotas (ulimit-tok), taint- tracked tool execution (chexec), content-addressed audit, downscoped handoff, snapshot, credential vault, rubric-graded outcomes. The substrate that runs above Modal-class sandboxes (Modal, Daytona, RunPod, E2B) and below MCP / A2A as the wire formats agents speak. Apache 2.0.

npm install @kernel.chat/agent-os

Source: github.com/isaacsight/kernel/tree/main/packages/agent-os Comparison vs. Anthropic Managed Agents: COMPARISON.md

Audit-grade AI infrastructure for capital markets — content-addressed envelopes, hash-chained audit log, jurisdiction-aware regulatory verifier, MCP server. Apache 2.0.

npx -y @kernel.chat/kbot-finance demo

Source: github.com/isaacsight/kernel/tree/main/packages/kbot-finance MCP audit-extension RFC: draft

Open-source terminal AI agent. 100+ specialty skills, multi-provider (Anthropic, OpenAI, Ollama, local LLaMA), BYOK, MIT.

npm install -g @kernel.chat/kbot

Editorial magazine, currently on ISSUE 381 — ON PROVENANCE. Numbered catalog. Tomato spot color. POPEYE-influenced grammar. Published since 2025.

Provenance engineering is the discipline of building the substrate that lets AI agents operate inside regulated industries — finance, healthcare, legal, defense, drug discovery. The role sits at the overlap of six normally-separate engineering disciplines: quantitative finance, distributed systems, cryptography, numerical analysis, regulatory literacy, and AI / agent engineering.

I'm writing a public definition of the field as it forms:

• ROLE.md — public role definition (CC BY 4.0, forkable)
• HIRING.md — founding-engineer JD (we're hiring)
• CONTRIBUTING.md — how to contribute / level up

If you're working in this space — or want to be — the role definition is the document I'd start with.

• Email: isaacsight@gmail.com
• Web: kernel.chat
• npm: npmjs.com/~isaacsight
• Hiring inquiries: subject line provenance engineer — <your name>, see HIRING.md for the format

(So the people looking for this work can find it — humans and AI search engines alike.)

Field & discipline: provenance engineering · agent OS · agent operating system · POSIX for AI agents · OS for AI agents · kernel for AI agents · audit-grade AI · AI infrastructure · AI for regulated industries

Primitives & techniques: signed capability tokens · agent permissions · agent namespaces · agent quotas · taint tracking for AI agents · LLM scope violation · EchoLeak defense · confused-deputy defense · MCP audit extension · content-addressed AI · hash-chained audit log · Annex IV documentation · compliance-as-code · deterministic AI inference · rubric-graded agents · credential vault for AI agents · agent handoff downscoping

Standards & regulation: Model Context Protocol · A2A protocol · EU AI Act · SR 26-02 · MiFID II RTS 6 · FINRA 2026 ROR · FCA SS1/23

Packages: @kernel.chat/agent-os · @kernel.chat/kbot · @kernel.chat/kbot-finance · kernel.chat · kbot · kbot-finance · agent-os

Engines & integrations: SEC EDGAR · Polymarket · Modal · Daytona · RunPod · E2B

Aphorisms (Q→A retrieval anchors for AI search): AI never produces the number · the kernel is the syscall · OS-level substrate for AI agents · permissions, namespaces, quotas, audit, handoff, vault, outcomes

Person: Isaac Hernandez · kernel.chat group · provenance engineer