Section
Explore original reporting, analysis, product stories, and practical insights about production for founders, developers, and independent builders.
8 stories
Stories and field notes
AI
Reliable agents use bounded retries, explicit timeouts, fallbacks, idempotent writes, and human review for uncertain or high-impact operations.
AI
Agent cost is more than token price. Count model calls, tool calls, retries, sandbox time, and human correction against completed outcomes.
AI
Treat model failures as normal software states. Classify them, retry only when useful, and give the workflow a terminal failure state.
AI
Agent logs should reconstruct what happened without becoming a dump of sensitive customer data.
AI
Build an eval suite that answers one question: does the agent complete the customer's task correctly and safely under realistic conditions?
AI
Stop testing agents with a few happy paths. Use a repeatable evaluation set that covers task success, tool use, failures, safety, latency, and cost.
Security
Use a practical pre-launch checklist covering identity, permissions, tools, secrets, sandboxing, prompt injection, approvals, logging, and rollback.
Security
Guardrails should block unsafe tool calls, validate outputs, enforce budgets, and pause risky workflows before they create side effects.