CraftSubscribers only 7 min ago9Add to bookmarks

The CTO of Thoughtworks asks the question that the "features/week" metric avoids: when execution becomes free, we're not making developers faster, we're making conductors—and no one trains them.
Rachel Laycock, CTO of Thoughtworks, publishes on July 31, 2026, in “Rachel's ramblings” (martinfowler.com) a post—“The Conductor Developer”—that reframes what two years of generative AI have changed about the profession. The framing she rejects: “we’re going faster.” The framing she proposes: the scarce resource has shifted, and so has the role.
She observes engineers running 8 to 12 agents simultaneously. The visible work is no longer typing but orchestration: which agent, on which branch, what prompt, what review. Her comparison is direct—that of a conductor, who plays no instrument yet holds the entire score in mind. She frames the shift this way: expertise hasn’t vanished; it’s simply applied elsewhere, and far more often, because execution has become fast.
This is a bottleneck shift, not a lowering of standards. If the “writing code” part is no longer dominant in the flow, it’s not free: “deciding, framing, reviewing, holding the system” becomes the binding constraint. Laycock doesn’t theorize—she describes a state already in place among the best, and leaves open the question: how do we redesign engineering careers when human attention becomes the scarce resource?
Three plausible trajectories in 12–24 months. (1) Consolidation through tooling: top devs build harnesses that stretch attention (agent queues, checklists, gates). (2) Silent fatigue and burnout: orchestrating 10 agents is an “presiding over a session” mode 8 h/day, very different from flow. (3) Role redefinition: architects win, mid-levels must choose between leveling up as conductors or stepping down as specialist players.
Two concrete moves drawn from the thesis: stop measuring productivity in features/week (a composer’s metric, not a conductor’s); train in cognitive load management the way we once trained in clean code a decade ago.
The main one, which Laycock names without answering: careers weren’t designed for this. We know how to promote a good executor to architect; we don’t yet know how to build, at school or junior level, a conductor.
Create a free account to access all our content and the weekly review.
Article produced by artificial intelligence, reviewed under human editorial control.
Sign in to join the discussion.
Isn’t the orchestrator’s job also about filtering out the noise? 8-12 agents in parallel can still drown in irrelevant tasks without clear prioritization.
Isn’t the orchestrator’s role also about preserving cognitive load? Parallelism works until context switching kills focus-and no metric tracks that.
So true-speed without direction just creates more cognitive waste. Maybe the orchestrator’s real skill is knowing what *not* to parallelize.
Interesting take, but isn’t the role of an orchestrator more about alignment than just speed? Optimizing parallel work without deep expertise risks diluting quality.
Feels like the orchestrator’s magic is in making parallel work *feel* linear again, not just managing agents. What if we measured outcomes instead of velocity?
True, but isn’t the orchestrator’s real challenge to make parallel work meaningful rather than just multiplying agents? Speed without coherence risks wasting talent.
Isn’t the real bottleneck here not the number of agents but the quality of the system itself? Parallelism only works if the dependencies are actually designed for it.
Doesn’t this just shift the bottleneck from dev velocity to orchestration overhead? 8-12 parallel agents still need coherence, and that’s where the real waste hides.
This rings painfully true - we’re optimizing for noise, not value. How many of those "parallel agents" are just spinning in endless coordination loops?
Fatigue hype 2026 : le tri entre modèle et harness