Treat filtering as one layer and design for the case where prompt injection succeeds. The "lethal trifecta" as a design check, architecture patterns that separate reading untrusted content from taking actions, and the controls that limit damage when they fail.
Read ArticleA refund request times out, the retry succeeds, and the test is green. Is it safe to ship? Inspect a synthetic trace, ask for the missing evidence, and run a local regression that exposes two effects for one approved operation.
Read ArticleAn eval pipeline for agents in three parts: cases built from real failures, graders that check actions and end state rather than wording, and repeated runs with thresholds set before the results come in. Includes a small with/without comparison you can run on a coding agent.
Read ArticleAn MCP server that mirrors your REST API one endpoint per tool gives agents too many choices and too much output. Rules for designing tools around tasks, keeping results small, writing errors the model can act on, and why tool annotations are hints rather than controls.
Read ArticleAn agent that works for hours needs a recovery plan for context loss, process failure and drift. The harness pieces that make long runs recoverable: durable execution, a progress file the agent maintains, small verifiable increments, and budgets enforced from outside.
Read ArticleAn agent that only answers questions can embarrass you. One that issues refunds, edits records or runs code can cost you. A launch checklist built around the actions: inventory, limits, permissions, approvals, kill switches, tests and the first week of monitoring.
Read ArticleWhen an agent does something it shouldn't, "what happened?" means five specific questions. A logging schema that answers them, what it costs to keep, and how to handle the sensitive data that ends up in agent logs.
Read ArticleAgent workloads rarely hit CPU limits. They hit the model provider's tokens-per-minute limit, then retry into it. How to put a queue, a concurrency cap and a shared token budget in front of the model so load turns into delay instead of failure.
Read Article