Essays and notes on governed AI in financial services. Essays argue one idea; notes are shorter takes on what I am reading and testing.
In June, Google Cloud published the Open Knowledge Format: knowledge as plain Markdown with a little structured metadata. It is the right foundation. Regulated enterprises need the next step: knowledge that states its own owner, provenance, freshness and trust.
Documentation fails because it is a second job. The knowledge we need is already produced as a by-product of work: commits, tickets, incidents, reviews and decisions. Capture it passively, distill it, and keep a human approval gate before anything becomes trusted.
AI has made writing code cheap. Verification, trusted context and human judgment are now the scarce resources, and teams that redesign their lifecycle around them will pull away from teams that only add assistants.
The tempting answer to fragmented context is to pour everything into one place. We tried that with data, and many lakes became swamps. A context layer has to be a curated overlay: retrieval-first, owned, and pointing back to the systems that hold the truth.
Most enterprise AI pilots hit the same wall: systems record what happened, but the reasoning behind decisions evaporates. Durable, governed context about the why is the real foundation for useful agents.
The knowledge an agent needs is scattered across code, tickets, chat, wikis, incident tools and people's heads. Each fragment is accurate and useless alone. Fragmentation, more than model quality, is what makes enterprise agents brittle.
A tour of aac.muthub.org, the open prototype I built to show architecture as a control plane, and the design choices that mattered most: catalogs, executable guardrails, waivers as data, a computed scorecard and an MCP server for coding agents.
Enterprises budget for compute, models and licenses, but almost none budget for the context those models depend on. Unmanaged context is the largest cost nobody tracks, and every new AI agent adds to it.
Handing an AI agent a production codebase without a harness is like putting a teenager in a sports car with no dual brakes. Deterministic guardrails and sensors, owned as a platform, are what make agents safe to scale.
AI now writes code faster than any review board can read it, while architecture still moves at meeting speed. Governance has to become something that executes in the pipeline and in the agent's context.
Shipping AI-generated code without recording where it came from is unmanaged risk. Regulated teams need deterministic provenance for non-deterministic models: line-level attribution, Git-native session history and AI bills of materials.
Platform-embedded, sandboxed and evolutionary coding agents each earn trust in a different way. Then the editor became a fleet manager. Choose and govern agents by how their work will be verified.
For complex work in regulated sectors, prompt-to-code cannot pass an audit. Making the specification the contract an AI must fulfil can, provided the spec stays anchored to the code and machines check conformance.
Data platforms built for people asking questions will strain when the main consumer is software taking actions. Five principles for designing data platforms where agents are first-class users, and why an open lakehouse helps.
As agents multiply, every enterprise needs an authoritative directory of the tools they can call. What a governed MCP registry must do, what the official registry leaves to you, and the mistakes to avoid.
In a large organization you rarely own the teams you need to move. Architects lead through clarity, credibility and standards people want to follow, and by treating the North Star and today's fires as one job.
Python is unrivaled for research and prototyping. Production enterprise AI is mostly integration, security and scale, and with JVM agent frameworks now maturing, the runtime that owns the transaction is often the right home for the agent.
Agents act on their own, at machine speed, across many systems. Static credentials and perimeter thinking don't hold up. Attested identity, explicit delegation and policy enforced outside the model do.
The Model Context Protocol standardizes how agents connect to tools and data. It is silent on whether they should. Four risk areas a regulated institution should resolve before agents get tools.
A trained model is the start of the work, not the end. The lessons of industrializing machine learning carry straight into generative and agentic AI, which is why AI engineering has to unify DataOps, MLOps and DevOps.
Earlier essays on data platforms, MLOps and cloud architecture, published on Medium.