Notes from the engine room.
Practical notes on system design, failure and recovery. Explore the engineering decisions through clear explanations and illustrative AI examples.
Reusable agent skills. Controlled execution.
What belongs in an agent skill, what the tools must enforce and how to review, test and operate reusable AI workflows.
Change the model. Recheck the work.
A release method for AI model changes: compare representative tasks, check outcomes and permissions, investigate failures and keep a route back.
When the sources disagree, show the disagreement.
Handle conflicting, outdated or inaccessible evidence with source checks, effective dates, access controls and a clear way to show what remains unresolved.
The agent stopped. Did the action happen?
Recover interrupted AI workflows by checking what happened, preserving the action record and deciding when to retry, seek approval or correct a completed action.
Prove parity before you rewrite
Capture a legacy system’s behaviour in tests before replacing it. Explore gradual migration, data reconciliation and where AI can help engineers.
What NetWare taught us about AI agent permissions
Lessons from NetWare effective rights for Microsoft 365 agents: trace each identity, permission and resource boundary to establish what the agent can reach.
RAID to retrieval: failure modes that never changed
Storage and AI retrieval need different integrity checks. Lessons from RAID about silent failures, controlled rebuilds and checking evidence end to end.
Capacity planning, from Hyper-V clusters to model routing
Apply the lessons of N+1 capacity, headroom and graceful degradation to model quotas, response time and the cost of completed AI tasks.
Least privilege for AI agents in Microsoft 365
Broad mailbox access increases the potential impact of prompt injection. Limit what an agent can read, change and send, and test those boundaries.
What a DR test should actually prove
A successful backup job is only the start. Test the data, dependencies, access, timing and people needed to restore a working service.