Working notes on responsible AI in production — agentic architectures, governance and risk controls, MCP and A2A, and the practical engineering of trust at enterprise scale.
In a multi-agent system, the risky moment isn't the model. It's the hand-off, where one agent takes another's answer and folds it into its own reasoning as settled fact, and the doubt, the source, and the authority behind that answer quietly fall off at the seam. Why agent-to-agent protocol is where identity, provenance, and the audit trail either survive the hop or get laundered into confidence nobody can defend.
Read article →An accurate model a clinician won't act on, or a payer can't defend, returns nothing. In medicine the value shows up only when someone signs their name under the decision. Why fairness, privacy, transparency, and human oversight aren't the compliance cost of healthcare AI, but the mechanism that turns an accurate model into a used one, and a pilot into a program that survives its first audit and appeal.
Read article →Changing a prompt, swapping a model, or handing an agent a new tool changes what your system does in production, yet none of it clears the gate a line of code has to. Why a versioned eval suite is the release control that turns responsible AI from a policy you intend into one you can prove, and why that gate belongs in the substrate.
Read article →Grounding ties an answer to a real document, but it never asks whether that document is still true. When a policy is revised or a supplier is suspended, the retired copy sits in the index and keeps answering as if nothing changed. Why lineage, freshness budgets, and retraction propagation are the controls that separate a grounded system from a merely confident one.
Read article →Sending every request to your best model feels like the safe choice. It's the expensive one, and often not even the better answer. Why routing each request to the smallest model that clears the bar is a cost control and a quality control at once, and why that decision belongs in the substrate, not in every app.
Read article →An agent in production is a user you never onboarded — no name, no scope, no end date, and usually a shared service account with more reach than the task needs. Why non-human identity, least privilege per task, and short-lived credentials are the governance controls that decide how far a compromised agent can travel.
Read article →Any model will hand you a confident answer in seconds. That fluency is free, and it's the trap. In an enterprise, an answer with no idea where it came from is a liability in a nice font. Why grounding, provenance, and the discipline to say "I don't have that" are the real controls behind trustworthy AI.
Read article →Software fails loudly. AI erodes quietly. A model that passed its acceptance test in March is not the system running in July, and a governance sign-off that ended at go-live is a photo of a system that already left. Why the real control is the evaluation that never stops running.
Read article →The pitch decks keep adding agents. Ours keep removing them. Why the number of agents in your architecture isn't a measure of how advanced it is, but of how much control you've quietly given up, and how it drives your risk and your bill.
Read article →The whole AI boom was aimed at one person: the developer. Meanwhile the claims analyst, the sourcing manager, and the investigator were still doing the work by hand. Why we build for the person making the decision, not the person shipping the code.
Read article →A compromised chatbot says something wrong. A compromised agent does something wrong, with your credentials, inside your systems. The security question for enterprise AI isn't whether the model can be tricked. It's how far the damage travels when it is.
Read article →The AI sticker shock that lands six months into production is almost never about the price per token. It's about the consumption pattern your architecture quietly committed to — long before anyone looked at an invoice.
Read article →Enterprises don't have an AI problem — they have an AI sprawl problem. Why we build the expensive, dangerous machinery once, and let governance, guardrails, and agents compound across every solution in the portfolio.
Read article →Most enterprise AI conversations argue about model choice. The harder questions are about what the model is allowed to touch, how it talks to other agents, and what that wiring leaves auditable. A field view of MCP and A2A in production.
Read article →A chatbot can be wrong. An agent can be wrong and act on it. Where autonomy actually breaks down in claims, investigations, and sourcing — and what human oversight has to look like if it isn't going to be a slogan.
Read article →Policy frameworks have outpaced production reality. The bottleneck in 2026 isn't principles — it's the runtime controls that make those principles enforceable when an agent is actually deciding something on your behalf.
Read article →