From Prompt to Action: Engineering Agentic AI Systems for Real World Deployment

About This Activity

A workshop on the distance between an agent that answers and an agent that does work, and on what has to be engineered to cross it. The session covered the harness rather than the model: tool interfaces designed so a wrong call is structurally impossible rather than merely discouraged, retry and backoff behaviour that does not amplify a fault across a fleet, human checkpoints placed where a mistake would be expensive, and observability that records why an agent chose an action rather than only what it did. Failure modes got as much time as capabilities, because the ones that matter in deployment are quiet: a loop that silently truncates its own context, a tool that returns success on a partial write, an agent that retries a non-idempotent operation. The closing argument was about write access. Read-only agents are a different risk class entirely, and most teams should ship several of those before giving anything permission to change state in a production system.