A production agent can give the right answer and still refund the same customer twice, act on a stale record or quietly finish half a job. Reliability therefore depends on the controls around the model: explicit state, idempotent tools, traces that show what changed and evals that test the whole run. Those controls also expose two less obvious problems. An expert's procedure has to become instructions an agent can follow, and a working system sometimes needs rebuilding before its accumulated fixes make it impossible to reason about. Cost belongs in the same picture. Token spend only means something when a team can connect it to completed work, failures avoided and outcomes users actually received.
Why first: Munaf changes the diagnosis. The moment an agent can call an API, it inherits the old distributed systems failures: a timeout with an unknown outcome, a retry that repeats a side effect, stale context. The other talks solve particular versions of the problem he lays out here.
Why second: Munaf describes the problem. Baraiya gives the first answer: let the LLM plan the next step and let deterministic harness code execute it. Approval, idempotency and the record of side effects such as a cluster restart or an email then live in that execution layer, not in whatever the model decides on the day.
Why here: Baraiya splits the brain from the hands. Milev and Kanagala show what Navan runs in production. They did not begin with multi-agent orchestration. They use one agent that loads skills on demand, plus hooks that capture goals and reasoning when ordinary logs disappear into model output.
Why here: the talks before this are about controls. This is about the input those controls act on. Buykin found at Maersk that screenshots of human clicks do not make an automatable procedure. An agent needs the preconditions, decisions, identifiers, checks, recovery steps and evidence written down. It is the least glamorous talk in the pack and probably the most useful.
Why here: you have built something by now. Kus warns that its half life may be only a few months. He tells an engineer to rebuild a working agentic search system just after it ships. His test for making that call is reassuringly dull: not a new paper, but an eval set showing better outcomes for customers.
Why here: Kus lets an eval set decide whether to rebuild. Grant asks what evidence earns an agent more authority at all. A pilot whose drafts match the analysts only proves the technology works. She wants a baseline before deployment, a named owner for each decision and autonomy raised in stages, and warns that a falling override rate can mean weaker review rather than a smarter agent.
Why last: Grant sets the baseline before the agent ships. Hong measures what the spend bought afterwards, treating token spend as a useful diagnostic and a terrible target. At Ironclad, the team moves from counting lines of code to open pull requests, merged pull requests and, finally, merged work weighted by complexity. Send this one to whoever asks what the spend bought.