Most agents call a hosted model API and need no GPU at all. When self-hosting a model does make sense, the decision comes down to memory, throughput and utilization. How to size the GPU and work out the break-even against API pricing with your own numbers.
Read ArticleThis site's publish pipeline pointed at an S3 bucket that never existed, and nobody noticed for ten months because an earlier step always timed out first. What that incident teaches about pipelines for agent deployments, where skipped checks can hide an untested deployment path.
Read ArticleAsk a coding agent for an orchestrator role that launches ECS tasks and you will often get ecs:* and an unconditioned iam:PassRole. Here is the escalation path when a privileged role trusts ECS, the scoped version, and a checklist for reviewing agent IAM before merge.
Read ArticleCoding agents produce Terraform that plans cleanly and hides its risk in a few lines. Where to look in the diff, and a Conftest policy that fails the plan when an IAM statement allows iam:PassRole without a service condition.
Read ArticleLambda is a good home for agents until a session outlives a single invocation. Where the limits actually bite (15 minutes, the API Gateway timeout, streaming), and when to use Step Functions, Lambda durable functions or a container instead.
Read ArticleAgents that write and run code need a sandbox, and a default Docker configuration is not sufficient for every threat model. What containers share with the host, when to reach for gVisor or a microVM, and a hardened docker run you can use today.
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 Article