Insights
Notes from the work.
Practical writing on building enterprise AI that reaches production.
Every AI agent has three identities. The worst failures happen in the one nobody named
Its own name, the person it acts for and the cloud role it runs as answer different questions. This week's Google, Databricks and Zenity news shows why each needs an owner and a test.
Read
Microsoft's support agent was accurate and still underused. Context, honesty and a way out fixed adoption
Microsoft's research on its Employee Self-Service Agent found people judge the whole support journey. Acknowledging what they already tried and a clear way to a human mattered most.
Read
An HR agent should not need a second copy of Workday. Federation moves the risk to the connection
Databricks' Workday federation, in beta, and Google's borderless lakehouse both read Workday data in place. That removes the copy and makes the integration user and query scope the real controls.
Read
Google's industry agents inherit ethical walls and cite sources. Industry AI is about rules, not vocabulary
Gemini for Financial Services and Legal, both in preview, lean on matter permissions from NetDocuments and iManage and on lineage. What that solves for regulated teams, and what it does not.
Read
Microsoft's own IT says to pick the simplest agent platform that works. Review effort should follow risk
A Customer Zero guide shows how Microsoft routes builders between Agent Builder, Copilot Studio and Foundry, automates reviews by risk color and treats maker friction as a security metric.
Read
Zenity's AgentCore research: the prompt was the way in, the agent's cloud role decided the damage
Zenity Labs reports that a prompt injection against one AgentCore agent reached other agents and their memory, because the default execution role spanned the region. AWS has since tightened it.
Read
Google's coworker agents get an email address, a calendar and a directory entry. Who offboards them?
At Gemini at Work, Google described agents with their own Workspace accounts and sub-agents that each get an identity. Before the features, enterprises need a lifecycle for agent identities.
Read
An agent can release 28,000 hours and save half the money promised. Hours are capacity, not value
Business cases for AI agents still lean on hours saved. This week's evidence shows why that number overstates value, where the larger value hides, and which costs it leaves out.
Read
Not every enterprise AI problem needs an agent. BigQuery's TimesFM 3.0 puts forecasting inside the database
Google added TimesFM 3.0 to BigQuery's AI.FORECAST in preview, with forecasts across many series and covariates. For forecasting, a model next to the data can beat an agent built around it.
Read
Mistral Large 4 is an API today and open weights by month's end. Use the gap to decide who runs it
Mistral released Large 4 in public preview, with 1 trillion parameters and 52 billion active, and promises downloadable weights by the end of October. The weeks between are a test window.
Read
Is the agent the product? Copilot Studio's October update suggests the product is the whole process
Microsoft now builds apps, deterministic workflows and agents in one Copilot Studio environment, and extends evaluation from agent conversations to automation. Approval has to follow.
Read
Agent guardrails are not one layer. They are six authorities, and no one should hold them all
Identity, model access, containment, sequence, change approval and spend answer different questions. This week's releases show them separating. A framework for deciding who holds each one.
Read
Have an AI use case stuck between prototype and production?
Tell us what you’re trying to ship. We’ll reply with honest next steps.
Discuss a use case