Running agents in parallel, and the coordination cost that eats the gain

Parallelism is the obvious appeal of multiple agents. Three problems, three workers, one third of the time.

The arithmetic ignores two costs. Briefing each agent takes your attention, sequentially. Reviewing three results takes your attention, sequentially. If the work itself was short, both costs exceed it.

So parallelism pays on long, independent tasks with compact results, and loses on short or entangled ones. The skill is telling which you have before starting.

Two tasks are independent when neither needs the other's result and neither modifies what the other reads. Both halves matter and the second is the one people miss.

Two agents editing the same file will conflict. Two agents editing different files that import from a shared module will produce changes that do not compose. Neither shows up until you review both.

Before splitting, list the files each task will touch. Overlap means sequential, not parallel — and if you cannot predict the files, that uncertainty is itself the answer.

The reliable pattern is several agents investigating in parallel and one agent — you — acting on what they return.

'Find every place this pattern appears', 'check whether this holds in the other module', 'summarise how this subsystem works' — three of these run concurrently without touching each other, and each returns something small.

Production work in parallel is where consistency breaks. Independent agents make independent decisions about naming, structure and approach, and three reasonable but different choices in one codebase is worse than one mediocre consistent choice.

An agent spawned mid-session knows nothing about the conventions established, the approach chosen or the alternatives already rejected. Anything you are carrying implicitly has to be stated.

For parallel work this means writing the shared constraints once and including them in every brief. Otherwise each agent invents its own answer to the same unasked question.

Also state the output format explicitly. Three results in three different shapes is more work to reconcile than the parallelism saved.

Every running agent needs briefing, tracking and reviewing, and all three happen in your head, one at a time. Past two or three, the tracking overhead dominates and you are managing rather than working.

There is a quality cost too: reviewing the fourth result carefully after three others is harder than reviewing the first, and things get approved that should not have been.

When you find yourself wanting many agents, the usual cause is that the task was not decomposed well. A better decomposition, run with fewer agents, generally finishes sooner.

The question captures both requirements at once. A task you could hand over in a paragraph is well specified and context-independent — exactly what parallelises.

A task where you would need to explain the last hour of decisions is one where the context is the work, and handing it over means transferring it badly.

Applied honestly, this rejects most candidates for parallelism and correctly identifies the ones worth it. The gain from those is real; the gain from the rest was never there.