Notes

How I think about agentic systems.

Short write-ups on the design decisions behind the systems on this site: what I chose, why, and where it can still fail.

• 7 Oct 2026|3 min read

  Fail-closed by default: MCP servers that agents can't talk their way past

  The model can ask; the server decides. What 'fail-closed' means in practice when AI agents touch business systems.

  MCPSecurityAgentsRead

• 7 Oct 2026|2 min read

  Making human approval survive a restart

  'Human in the loop' is easy to say. The hard part is the loop: what happens to a paused agent when the server restarts?

  LangGraphHuman-in-the-loopReliabilityRead

• 7 Oct 2026|2 min read

  Let the code do the maths and the model do the words

  How to keep LLM-written explanations honest: compute the facts in code, hand the model a small facts packet, then check what it wrote.

  LLMsGroundingEvaluationRead