Insights
Notes from the work.
Practical writing on building enterprise AI that reaches production.
IBM gives each AI agent an owner record and credentials that expire. The revocable credential is the control
IBM Agent Identity and HashiCorp Vault 2.2, both generally available, pair a record of who owns an agent with short-lived, scoped credentials that can be revoked mid-session. How the layers fit.
Read
Anthropic's drug development results with six partners are not one kind of evidence. Read them by type
Claude shortened a CAR T construct by about 70% in lab tests and cut a manufacturing document review by 89%. Lab results, computational output and workflow gains need different follow-up questions.
Read
Gemini 4 Carbon goes to cyber defenders first. The access gate is part of the product
Google's new model leads on vulnerability finding and long-running work, but only selected defenders can use it for now. What the benchmarks show, what they don't, and how to prepare.
Read
Oracle and ServiceNow now let employees work from ChatGPT. Both keep the authority to act for themselves
Oracle's Fusion Companion and ServiceNow's ChatGPT plugin move the conversation to OpenAI while approvals, permissions and the record stay in the system of record. Design the join between them.
Read
Anthropic's AI built full exploit chains in Chrome, Windows and Linux. Triage by single CVE no longer fits
An unreleased Claude model found and chained vulnerabilities in five major platforms, including 62 Chrome CVEs. The lesson for security teams is about how flaws combine, not only how many get found.
Read
Before an agent runs a system, it should be able to explain it. Most systems don't keep enough history
The most useful first agent in a complex system is often read-only: it explains what happened. That depends on four records most platforms keep badly, and finding the gaps is the first deliverable.
Read
Ollama now speaks, locally. Only its English model allows commercial use, so read the license first
Ollama added text-to-speech with Gradium's Phonon models and an OpenAI-compatible API. It runs on Apple Silicon only for now, and only the English model allows commercial use.
Read
GitHub now counts Copilot tokens used from Claude Code, Codex and OpenCode. Use it for cost, not productivity
Copilot's usage metrics API now splits model requests, sessions and tokens by client, including tools outside GitHub's own plugins. Useful for cost and adoption, misleading as productivity.
Read
Four kilometers underground, there is no cloud. Google's TB screening AI took seven years to get there
Google Research traced how its chest X-ray models became a CE-certified screening tool used in about 40 facilities, including offline in a South African gold mine. The long part was never the model.
Read
Your AI gateway holds keys to every model and tool your agents use. Protect it like identity infrastructure
LiteLLM's compromise and eight fixes in the official MCP servers point one way: the parts between agents and systems now concentrate credentials. A framework for treating them as tier zero.
Read
Bayer's 22 agents don't run its plant-breeding models. They explain what happened inside them
Bayer Crop Science built a read-only agent layer on Amazon Bedrock AgentCore over a platform running 500+ ML model configurations, and reports 80% of questions answered without an engineer.
Read
Microsoft wants every agent to learn your business once. Work IQ puts definitions and procedures in Dataverse
Microsoft is previewing semantic models, versioned Business Skills and governed actions in Dataverse, open to Copilot, Cowork, MCP clients and n8n. Business context becomes shared infrastructure.
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