Journal

Notes on agentic development.

How to turn AI into real shipping speed: audits, workflows, modernization, and the tactics that hold up in production.

We killed an agent pipeline and a microservice before launch—because the real failure was state, not compute.

After an enterprise launch cratered organic-to-chat accuracy, we built answer-engine routing from llms.txt instead of guessing.

We had an LLM hallucinate a compliance-friendly answer; the fix wasn’t better prompts—it was guardrails, audit trails, and a real kill-switch.

We once shipped a “safety” kill-switch that blocked responses—but still logged regulated data. That was the real failure.

We shipped an LLM agent for a regulated workflow; the first failure wasn’t accuracy, it was missing audit evidence.

We killed an agent in production, not because the model was wrong, but because the workflow lacked auditability and hard failsafes.

Page 3 of 45

Previous 2 3 4 5 … 45 Next