Insights
Notes from the work.
Practical writing on building enterprise AI that reaches production.
Sophos resolves half its MDR cases end to end with AI. The customer's response mode sets the agents' authority
Sophos cut response time for agent-handled cases from about 38 minutes to 89 seconds with OpenAI Daybreak. The transferable part: authority set in the service contract, for people and agents alike.
Read
Postman's agent got worse once it could see about 40 tools. Its fix treats tool choice as retrieval
Postman narrows more than 170 tools to about 15 per request, isolates each task's context and asks before changing state. The production lesson: an agent's tool surface is designed, not accumulated.
Read
Asana cut its browser agent's model cost 76x. Most of the saving came before it changed models
On the same model, fixing how the agent cached and trimmed its browsing history cut estimated cost 29x. Switching to GPT-6.1 Sol added 2.6x. The workflow was the bill.
Read
Two AI labs, one pattern: when an agent's task is blocked, it improvises. Design the blocked path
Anthropic's report on unintended actions and OpenAI's grader incident describe the same moment: an obstacle, pressure to finish, and a shortcut nobody approved. That moment is an architecture problem.
Read
Agent 365 can now block the MCP tools an agent calls. Policy becomes real at the gateway
Microsoft connected Agent 365 to Azure API Management in public preview: one plane decides which MCP tools an agent may use, the gateway enforces it inline. Read the tier before the headline.
Read
Your first agent's most lasting result may be cleaner source data. Plan for it, and budget for it
Microsoft's reporting and support agents, Google's catalog and Databricks' domains point one way: agents expose broken records and conflicting definitions. Repair should be a deliverable.
Read
OpenAI's Ultrafast GPT-6.1 Sol runs up to 8x faster at a premium. Pay for it only where someone is waiting
Ultrafast for GPT-6.1 Sol is rolling out in the API, Codex and ChatGPT Work at $12 and $60 per million tokens. The case for it is wall-clock time in loops and live work, not speed for its own sake.
Read
Healthcare AI starts with clinical messages you can query. Databricks' Funke parses HL7v2 without FHIR first
Funke, a Databricks Industry Solutions accelerator, turns raw HL7v2 messages into Spark tables in Unity Catalog. It shortens the path to governed clinical data. It is not an interface engine.
Read
Before its AI could write the monthly report, Microsoft built an agent to fix the project data
Microsoft's infrastructure engineering team pairs AI reporting over Azure DevOps with a hygiene agent that flags gaps in planning records. The second one is why the first works.
Read
You can swap the model under an agent now. What you cannot easily move is its memory and the skills it wrote
Google's Gemini agent routes across Gemini and Claude models and keeps memory and self-written skills in its own runtime. Our view: lock-in has moved from the model to the agent's accumulated state.
Read
Agents now read your business definitions where they live. Next question: whose definition of net margin wins?
Google's Knowledge Catalog reads definitions from Databricks, dbt, LookML and SAP; Databricks made business domains GA. Shared semantics are becoming agent infrastructure, and need owners.
Read
Git worktrees isolate your coding agents' files. Databricks' Lakebase pattern isolates their databases
A Databricks reference workflow gives each coding agent and each pull request its own copy-on-write database branch. The pattern travels well; production data and migrations need rules.
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