Insights
Notes from the work.
Practical writing on building enterprise AI that reaches production.
Stop asking which jobs AI will replace. Workday's Work Intelligence measures the tasks that change
Workday's early-access tool breaks jobs into tasks, links them to skills and tracks how work shifts with AI. The task is the right unit for planning. The risk is treating its output as decisions.
Read
Anthropic ran a small AI model on $55 TV boxes. The enterprise lesson is about devices you already own
Anthropic's red team found small open models can set up their own inference on cheap connected hardware and, in simplified tests, replicate. The defenses are old ones: inventory, defaults, traffic.
Read
The HR agent nobody prompted: what changes when Workday's agents work in the background
Workday's Autopilot agents watch recruiting, payroll and accounting and act before anyone asks. Only recruiting is live, for early adopters. The question: how often they look and what they may do.
Read
Agent skills are moving to a shared library over MCP. A fingerprint proves the file, not the advice
MCP's Skills extension is final, with Foundry and Databricks support in preview and beta. A shared library ends drift between copies, and makes one source instruct many agents.
Read
Chat windows are becoming interchangeable. Enterprise AI's real contest is who authorizes the transaction
Workday, SAP and Microsoft draw the same line: any assistant can start the work and any model can reason, but authorization and execution stay with the system that owns the record.
Read
A security policy that never reaches the coding agent is not a control. GitHub now lets Copilot fail closed
A new setting makes Copilot clients fetch current policy at startup and refuse to run if they cannot; the sandbox setting now forces old clients to update. The trade is availability for certainty.
Read
Your agent's score can change overnight while the agent stays the same. Snowflake's judge switch shows why
Snowflake is moving Cortex evaluations to new metric versions and judge models as the old judge retires on October 14. Scores across versions are not comparable; unpinned thresholds will move.
Read
AI agents need an offboarding process. Workday's deal for C1.ai treats it as identity governance
Workday agreed to buy C1.ai to tie agent access to the same lifecycle that governs employees. The deal has not closed, but the problem it names is already in every enterprise running agents.
Read
An agent proposed six overnight changes to a 205,000-line codebase. Engineers kept two. That is the result
Google ran Gemini 4 Argon unattended for 11 hours on its Agent Development Kit. Two changes shipped, four were rightly rejected, and a first run without repository rules broke a public schema.
Read
Microsoft's network agent has an unusual goal: turn its own successful reasoning into ordinary software
An Azure Networking engineer describes an on-call agent bounded by identity, typed tools and approval gates, whose repeated fixes become deterministic playbooks. That fourth boundary is the new idea.
Read
Your agent may read the document. Can the employee asking? Microsoft's SharePoint pattern answers per user
A Microsoft reference design keeps a Foundry hosted agent on its own identity for inference and runs SharePoint retrieval as the signed-in user. Its two-user test belongs in every RAG rollout.
Read
Workday goes headless for AI agents. Its bet is that the business rules travel with every request
Workday opened HR and finance to outside agents through MCP tools and live SQL, and is taking Sana into Claude, Copilot and Slack. The interface becomes optional; permissions and audit do not.
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