One Dish Versus a Full Service
the real difference between a cook who makes exactly one dish when asked and a kitchen that runs an entire dinner service — planning, delegating, checking its own work.
Part 1
Part 2
What the Expediter Actually Runs
the real components behind a working agentic system: a goal, a plan, tools, and a loop that keeps checking whether the goal has actually been reached.
Part 3
Before the Kitchen Could Run Itself
how multi-step automated tasks got done before agentic AI existed, and what that history reveals about what's genuinely new here.
Part 4
Reading the Ticket
how an agent interprets a high-level goal into something concrete and actionable, and why getting this first step wrong dooms everything that follows.
Part 5
Breaking the Order Into Stations
how an agent turns a correctly understood goal into a concrete sequence of actionable steps — and why this planning step is where most agentic failures actually originate.
Part 6
The Line Cook's Tools
how tool use and function calling let an agent actually take real actions in the world, not just generate text about what it would do.
Part 7
Calling for Backup
how an agent delegates a subtask to a specialized sub-agent, rather than trying to handle every kind of work itself.
Part 8
Tasting Before It Goes Out
why an agent that checks its own work before finishing produces meaningfully more reliable results than one that simply acts and stops.
Part 9
The Pass
why a dedicated checkpoint reviewing an agent's overall output, separate from the agent's own self-checks, catches mistakes individual self-verification misses.
Part 10
86'ing an Item
how a well-designed agent gracefully handles a genuine constraint — a missing resource, an unavailable tool — rather than pretending the constraint doesn't exist.
Part 11
Firing Multiple Tickets at Once
how an agent can run independent subtasks in parallel rather than working through a plan strictly one step at a time — and when that parallelism genuinely helps.
Part 12
When an Order Comes Back
how a well-designed agent handles a failed step gracefully — retrying, adapting, or escalating — rather than simply crashing or silently continuing with bad results.
Part 13
The Expediter's Memory of the Whole Service
how an agent keeps track of everything that's happened so far in a multi-step task, and why this memory has to be managed deliberately, not just accumulated endlessly.
Part 14
Calling the Manager
why the most reliable agentic systems know exactly when to stop and escalate to a human, rather than pushing through every decision autonomously.
Part 15
A Recipe vs. Improvising Off the Cuff
the difference between an agent that plans every step upfront and one that decides its next move only after seeing the result of the last one.
Part 16
Multiple Cooks, One Kitchen
why some tasks are better handled by several coordinating agents with distinct roles than by one agent trying to do everything.
Part 17
When the Kitchen Catches Fire
the guardrails that keep an agent's failure contained, instead of letting one bad decision cascade into a runaway loop or a genuinely costly mistake.
Part 18
The Health Inspector
why evaluating an agentic system requires testing its full, multi-step behavior, not just checking whether the final dish looks right.
Part 19
Running the Kitchen Every Night
what changes when an agentic system moves from a single successful test run into something that has to operate reliably, night after night, in production.
Part 20
The Whole Kitchen in Full Service
reassembling every station covered across this series into the picture of one complete, coordinated agentic system running a full service.