From Coder to Orchestrator

As agents do the typing, the engineer's job shifts from authoring code to directing it -- scoping work, choosing approaches, reviewing output, and steering several agents at once. The unit of work moves up a level: from writing functions to specifying intent and judging results.

The Pattern

"The developer is now the orchestra conductor." -- Rich Miller (source)

As agents do the typing, the engineer's job shifts from authoring code to directing it -- scoping the work, choosing the approach, reviewing the output, and steering several agents at once. The unit of work moves up a level: from writing functions to specifying intent and judging results. Less time in the editor, more on the ends of the job -- specification, design, and review.

Addy Osmani frames this as a move "from micro-manager to macro-manager," and draws a useful distinction between two modes. A conductor works closely with a single agent in a synchronous, interactive session -- the IDE or CLI "pair programmer," steering each step and approving each diff. An orchestrator sets high-level goals, assigns tasks to multiple autonomous agents, and lets them work asynchronously -- often returning ready-to-review pull requests. The two modes co-exist on a spectrum, and the same person flips between them through the day. What changes is the position in the loop: as Osmani puts it, "the human's 'coding' moves up a level to writing specifications and prompts." This is the orchestration end of loop engineering, and it depends heavily on the coding interfaces that make running a fleet practical.

Why It Matters

This is the spine of the whole theme. Every other shift here -- comprehension, learning, hiring, performance -- follows from the day reorganizing around direction rather than authorship. It is not people-management; it is orchestration of agents, closer to an engineering lead's judgment than a typist's throughput.

The leverage is real and increasingly concrete. Robert Brennan of OpenHands describes one client fielding thousands of new CVE announcements a day across a massive codebase: rather than work through them iteratively, "we send out agents in parallel to solve CVEs across the entire codebase -- 90% of them get resolved automatically," leaving humans to focus on the 10% that didn't make it through (a vendor account, self-reported and directional). The orchestrator's effort profile reshapes accordingly -- front-loaded into a clear task spec and back-loaded into review and integration, with little in the middle -- which is what lets one person carry more parallel work than single-agent pairing ever allowed.

The honest tension is that orchestrating at scale means you are no longer eyeballing every change as it is made. Osmani is blunt that human oversight "remains critical as the final failsafe," that current tools still require a human to review every PR before merge, and -- in his own aside -- that "the code review loop... is going to need work if all this code is not to be slop." The failure mode is orchestrating work you no longer understand, which is why this pairs tightly with the discipline of comprehension debt and leans on automated review and verification to keep the review loop from becoming the new bottleneck. The skill being rewarded is less typing and more decomposition, clear specification, and judgment about when to trust the agent versus drop back into conductor mode and fix it yourself.

Last reviewed: 2026-06-25