The Orchestrator

Chapter 1 argued that a new job is forming. This chapter is about the people already doing it — usually without a title, without approval, and without anybody to compare notes with. It follows Maya, a senior engineer who runs her own fleet of agents on her laptop, through the same returns problem from Chapter 1.

At 9:12 on Monday, Maya has three private agent workflows running on her laptop and one pull request open in the checkout repository.

The first agent reads the customer-service export and groups complaints by the problem customers are actually describing rather than by the tags the support team happened to apply. The second agent turns the resulting cases into regression tests. The third agent reads the Jira backlog, compares issues against recent support volume and incident history, and proposes what the team should look at next. None of these systems came through the company’s architecture review, and Maya built them because she was tired of waiting three weeks of meetings for an answer she could get in a few minutes.

She is a senior engineer at a company whose website sells home goods, and her title is still “platform engineer.” If you have read the Introduction’s note on method, you already know what Maya is: a composite, assembled from practices that exist today, not a person at a company you could visit. The practices are the evidence. She is the delivery. She is not writing code faster. In fact, she is not even writing that much code, because as an orchestrator she is directing the system to write code for her and then reviewing it at times. She is connecting customer evidence, product decisions, repositories, tests, and release signals into a working system that can decide what deserves attention next.

She is also doing all of this with tools nobody has approved. She runs a fleet of Hermes agents on her laptop and has worked out how to do it without attracting the attention of Infosec, and each agent carries a curated set of skills and a memory system she has been cultivating for months — systems she uses to orchestrate both how the product gets developed and how it responds to customers.

At 9:18, a product manager messages her: “How did you already know this was the returns API? We haven’t even had the support meeting yet.”

Maya has already read the conversations, or rather her systems have. They have found the main pattern, separated the cases that need different treatment, and shown her where the real problem is.

Her answer to the product manager is simple: “I let the agents do the work.” The product manager still doesn’t know what that means. He just says thanks.