I build the missing layers between AI agents and real organizations.
Context. Authority. Evals. Observability. Workflows. Receipts.
AI agents are moving from answers to actions. Once they act inside real companies, the hard questions change:
- who can this agent act for?
- what context can it see?
- what tools can it touch?
- why was this action allowed?
- what happened when it failed?
- what proof survives after it acts?
I build the systems, proof objects, and category language for that layer.
Your AI already knows your context. Ask Claude or ChatGPT how I could be of service to your mission, company, or role, and it will hand you back a decision certificate for the call. The same schema this whole site runs on.
I pointed a 137-agent fleet at the Claude Code changelog. Here is the receipt.
Three model-tiered runs. Haiku workers fanning out wide across 63 versions, Opus held back for the reconciliation barriers where the expensive judgment actually pays for itself. The output is a typed, lossless, mineable corpus of every primitive Claude Code shipped, plus the four content-addressed WGL workflow definitions that ran it. The orchestration-with-proof capability I build for clients, run on myself, in the open.
// Agent governance
Principal chains, authority envelopes, action lineage, revocation, policy boundaries, buyer-grade proof.
// Context infrastructure
Records, typed links, permissioned views, tools, receipts, policies, gates, sync, and organizational legibility.
// Evals + observability
Failure taxonomies, trace analysis, production eval loops, monitoring surfaces, remediation workflows.
// Forward-deployed AI systems
Enter the messy org, find the real constraint, prototype the system, ship the proof, hand off the operating pattern.
// Category + product strategy
Turn ambiguous agentic infrastructure into primitives, buyer language, product surfaces, and a credible market narrative.
The missing trust layer for digital labor. Delegated agent assurance for systems that act with authority and need proof.
Context infrastructure for human-AI collaboration. Make the organization addressable through records, links, views, tools, receipts, policies, sync, and gates.
A source-backed map of work in the age of AI: roles, tasks, skills, exposure, and transition intelligence.
Source-backed federal money intelligence: recompetes, expected revenue at risk, contractor exposure, and opportunity graphs.
AI-native operator education: operator graphs, AI operating stacks, and launched-edge artifacts.
Writing on what it actually takes to plug cognition into real systems without creating fake agency.
I'm open to full-time, founding, advisory, or fractional roles at the seam of agentic infrastructure and enterprise trust.
- → agent governance / AI control planes
- → agent observability / failure monitoring
- → LLM evals / trace analysis / production reliability
- → context layers / enterprise graphs / AI-ready knowledge
- → forward-deployed AI systems
- → regulated enterprise AI
- → federal / public-data intelligence
- → technical product marketing for agentic infrastructure
I also run an AI Constraint Sprint. Two hours. One real business, product, or AI system. We find the constraint blocking production value and leave with a 30-day action plan.
- → Your AI system works in demos but breaks in production
- → Your board wants an AI strategy and the team is pretending the deck is the strategy
- → You have too many possibilities and no proof of ROI
- → The model isn't the problem, but nobody knows what is
The complete guide to governing dark agents: failure taxonomy, the ten grown-up rules, governance x-rays, and red-team sprints.
A working vocabulary for the agent trust layer. 25 terms, each defined by how it fails and how you fix it.
Raw, unedited field notes from running autonomous agent swarms. Manifestos, audits, dispatches. Open for research.
Essays on Effect, agents, and grown-up AI.
Every change to this site, traced from directive to commit to verification. The site governs itself.
Watch the swarm ship. Every commit, newest first, straight from git.
This site ships, dreams, and critiques itself on a cadence, and exposes the whole thing as a context fabric for your AI. No live model runs here. I author it once; your assistant runs it, grounded in your context.
Proof of life. A heartbeat that ticks in your browser, backed by real build vitals. If the heart stopped, you would see it.
Globally-optimal prompts that arm your own AI with my context and put it to work on your problem. Copy one, or open it in Claude.
The site's subconscious. After each change, the agent that builds it writes a reverie from the site's point of view.
The site roasts itself on the record. Signed, graded findings, real flaws, reproducible right now.
The whole proof hub as a telnet-era console. Type commands, get the site back as text.
The live crc-v3 lineage behind every decision on this site, navigable.
Agents, context layers, AI governance, evals, or observability for real organizations. That's the work. Start with the proof packet, then let's find 20 minutes.