GLEN KEANE ON THE PROBLEM OF AGENTIC DRIFT

Dr. Strangeagent or: How I learned to stop prompting and love the State Machine, by Glen Keane.

In today's Engineering Takeover we describe what our man revealed when he took to the stage at AWS User Group Dublin last week. His core thought was a provocation: what if the way we're building AI agents has it somewhat backwards?

The talk centred on the problem of agentic drift. LLMs don't have built-in program trackers, call stacks or step counters. Give them a long set of instructions and they'll lose their place. Add more instructions to compensate and you compound the problem - more tokens, more cost, more unpredictable behaviour.

The natural instinct is to write a better prompt. Glen's argument is that the prompt was never the right tool for the job.

The alternative: let state machines do the driving. Remove some of the digital intelligence and replace it with an orchestration layer to handle sequencing, error recovery, retries and branching. The LLM gets called at specific points to do what it does well - reasoning, summarising, extracting meaning. Everything else stays deterministic.

The end-up? You get auditability, cost control and reproducible behaviour. The same mental model engineers already use for workflow orchestration, applied to a different runtime.

Thanks to Amazon Web Services (AWS) User Group Meeting Dublin ☘️ for the opportunity to collaborate and for continuing to bring the Dublin tech community together on cold Tuesday nights. These meetups are where practitioners share hard-won lessons rather than marketing slides. Long may they continue.

The first slides of Glen's talk are below, or you can see all of them here.