The Orchestrator

The ideas in this book take time. You cannot snap your fingers tomorrow, fire up a fleet of Pi or Hermes agents, and command legions toward perfectly managed goals — that is not happening — and if you pretend otherwise, treating orchestration as a weekend install rather than a practice that has to be learned, staffed, and governed, you get the Klarna shape: a public claim about capacity followed by a public admission that judgment was never designed.

Premature overadoption is commitment ahead of comprehension. A pilot of ten people succeeds, so the organization entitles everyone. A vendor demo looks inevitable, so the capital commitment arrives ahead of the unit economics. A founder hears Karpathy say he has not typed a line of code since December and concludes the company can do the same by Monday. Useful experimentation collapses into status theater, and being seen adopting AI quietly replaces asking what problem is being solved, what failure would look like, and what cost envelope success implies.

There is a human version of this that never shows up in the architecture diagram. A competent engineer who was proud of careful work starts to feel slow, because the room is celebrating the person who demoed six agents before lunch and nobody is asking whether those six agents share a permission model, an evaluation set, or a person who will be awake when one of them spends money. The careful engineer either joins the race or becomes the person described as “not leaning in.” That social pressure is part of the antipattern. Premature overadoption does not only buy the wrong tools; it teaches the wrong courage.

The orchestrator’s counter is almost boring, which is exactly why organizations skip it. Bound the bet, name what the pilot is allowed to prove, and separate “this worked for ten people” from “this is ready for the company.” Keep a written list of what the system may not do yet — and if you are already wiring tools, put the may not in a fail-closed policy, an ACS-style manifest or a host allowlist or a deploy gate, rather than only in a slide titled Guardrails. Premature overadoption hates lists, because lists make the fantasy negotiable. “We use MCP” is not a list of what the agent may decide.