Bits of enterprise AI, and what lies beyond the build.
Governance, agents, cost and the leadership calls behind AI in regulated enterprises. By Vivek Mathew.
Latest
The Approval Gate Finally Has a Place to Live: AgentCore Lifecycle Hooks
For months this series has said the approval gate for consequential agent actions must live outside the agent's reach, and that it was still yours to build. AgentCore harness lifecycle hooks are the first...
Read the post →Seven Architecture Decisions That Outlive the Model
Models change every quarter. These seven decisions don't: who may write to shared agent memory, who owns the knowledge base, how many false positives a guardrail...
Who Did It? Identity and Least Privilege for AI Agents
When an AI agent reads a record, sends an email, or moves money, whose permissions was it using? Most enterprise agents can't answer that, and the...
Four Kinds of Repeated Tokens, and Four Ways to Stop Paying for Them
Most of the tokens on most enterprise model calls have been sent before, by someone else. Sharing that context across users is four different problems with...
The Agent Harness, Explained: What It Is, and Which Parts Bedrock and AgentCore Give You
The model is the engine; the harness is the rest of the car. Most enterprise agent failures are harness failures. A component-by-component map of what a...
Bedrock AgentCore in Practice: What the Managed Primitives Buy You, and What They Don’t
AgentCore is a set of managed building blocks for agents, not an agent platform that thinks for you. A primitive-by-primitive look at what each one actually...
How to Say No to an Executive’s AI Demo Without Losing the Executive
A senior leader saw something at a conference and wants it running by the quarterly review. Refuse and you're the blocker; build it on the side...
Design Reviews Are the Best Teaching Tool You Have
Most engineers learn architecture not from courses but from having their designs reviewed. A review can be a gate, or it can be the most efficient...
Prompt Routing: Why Every Request Shouldn’t Go to Your Most Expensive Model
Most enterprise AI traffic is simple, and most of it is sent to a frontier model anyway. What prompt routing is, the three ways to do...
How to Talk to a Risk Committee About AI: Controls as Evidence, Not Friction
Most AI initiatives in regulated enterprises don't fail on the technology. They fail in the room where risk partners decide whether to let them ship. What...
MCP Servers in the Enterprise: What Changes When Tools Can Act
A chatbot that says the wrong thing is a content problem. An agent that does the wrong thing is an incident. What changes in your threat...
Observability for LLM Calls: What a Trace Should Contain, and How to Debug a Blocked Prompt
A guardrail that blocks a prompt with no trace is a black box with a 'no' button. What an LLM trace needs to contain, how to...
Cost Attribution for LLM Usage: What to Tag, What to Report, Who Reads It
LLM spend is invisible until you make it visible, and 'we spent $X on AI' is not attribution. A practical tagging model, the three reports worth...
Guardrail Enforcement Must Live in Your SCPs, Not as an Afterthought in Your Application Code
A guardrail a developer can forget to call is a control you will eventually lose. In an enterprise, enforcing LLM guardrails at the Service Control Policy...
AI Governance Is the New Cloud Security
Cloud security didn't get safe by asking developers to be careful. It got safe when the rules moved into policy nobody could bypass. AI governance is...