I work on the control layer for enterprise AI agents, and on the layers underneath that make the controls work.
Thirty years building systems where a mistake is expensive: federal health, defense, public health, finance. Currently technical lead for an AI specialist architecture team. I write specifications, build reference implementations, and ship the platform pieces that governed agents actually need.
wesjackson.ai · LinkedIn · Medium
Enterprise AI is segmenting into five layers. Control is the one I work on most, but it only works if the other four are there, so I work on those too.
| Layer | The question it answers | What I've built |
|---|---|---|
| Control | now that they can act, how do we stop them breaking something | ptc-gal-standards · ptc-gal-reference · stigcode · veripak |
| Context | make the AI understand my organization | memory-hub · retrieval-hub |
| Action | don't give me a copilot, do the work | fips-agents · mcp-test-mcp |
| Capacity | economical AI compute I can actually operate | upstream llm-d and vllm · workshop-setup |
| Economics | is this worth what I'm paying for it | platform-billing-operator |
ptc-gal-standards Two specifications for agentic systems, published under the Community Specification License 1.0. Provenance & Trust Context (43 conformance clauses): a signed trust-context object that travels with data across every agent and tool boundary. Grant & Autonomy Lifecycle (35 clauses): authority as stored, signed state per principal and action class, raised only through a maker-checker ceremony and lowered automatically by deterministic triggers. Both are filed for Linux Foundation agent-standards discussion; I sit on the committee.
ptc-gal-reference The implementation that shows how PTC and GAL can work. A deterministic tool broker that an agent cannot talk past. Controls that live inside the agent harness are advice to the component under attack; anything a model can be talked out of is not a control.
memory-hub Persistent, governed agent memory with scoped access, version history and audit. Built on the premise that memory is a context-assembly policy problem rather than a retrieval problem, and that the design should work backwards from the forensic question: what did this agent know when it acted, who wrote it, and was anyone allowed to see it?
platform-billing-operator Metering and FinOps for platform services on OpenShift. Token metering is not AI FinOps; the real problem is giving autonomous systems economic authority and being able to say afterward what it cost.
This work is spread across three accounts, for boring historical reasons.
- @rdwj — here. Platform components, MCP servers, developer tooling, most of the day-to-day.
- @wjatx — the standards work: the specifications, their reference implementation, and the Trust Bricks diagrams.
- @fips-agents — building agents that run where compliance is not optional: templates, a CLI, a gateway, and a code sandbox with AST guardrails, Landlock and seccomp isolation.