Everything Is Called an Agent Now. Almost Nothing Needs to Be One.

Ask a team what they are building and you hear "an agent". Ask which steps it takes and most of them can list those steps, in order, on a whiteboard, before any request has arrived. That is a workflow. It is cheaper and it pages you less. The difference is who decides. An agent decides what to run. A workflow runs the steps you fixed in advance. That one line tells you whether you can forecast the bill, cap the latency, and write the runbook before the incident. This session walks the five shapes these systems are built from, and what each one costs. Same support ticket through all of them. The dearest costs six times the cheapest for the same answer, and nobody picks that on purpose. The result worth staying for: parallelisation was slower than chaining. Three branches at once, then a merge call that has to wait for all three. At three branches the merge dominates. Concurrency buys latency, never tokens, and if your tracing adds up call durations you cannot see the benefit you paid for. Then the defect this shape usually ships with. An uncapped fan-out is a $2 request waiting to happen. What you will learn: - Why parallelisation can finish behind the code it was meant to beat - The one question that decides between all five shapes, and it is not a capability question - Chaining, routing, parallelisation, orchestrator-workers and evaluator-optimiser, with the fit for each - Serial time is the bill, wall time is the wait, and most tracing only shows you the first - Why an uncapped fan-out turns a $0.002 request into a $2 one - Why you do not have evaluator-optimiser until you can write the bar down - Why money, not capability, is the thing that should flip the answer