Skip to main content

How much faster is fanning out?

Splitting a task across subagents feels like it should make it proportionally faster at about the same cost. It doesn’t. Serial work caps the speedup, integrating each worker’s report is serial too, and every worker reads the shared context for itself.

The answer

4 workers: 1.33× faster, 1.53× the cost

45 min instead of 60, and 608K tokens instead of 370K. Each worker reads the 50K of shared context for itself, so it’s paid 4 times. Fastest at 3 or 4 workers (45 min). However many workers you add, it’s never faster than 2.5×.

  • With integration (solid)
  • Ideal Amdahl curve (dotted)
  • Ceiling at 1 ÷ serial fraction (dashed)

Hover or use the arrow keys to read a worker count. Click it, or press Enter, to use it.

The curve as a table
Time, speedup, tokens, and cost for 1 to 32 workers.
WorkersIdealWith integrationTimeTokensCost
11×1×60 min355K$0.75
21.43×1.25×48 min454K$0.99
3 (best)1.67×1.33×45 min531K$1.18
4 (current)1.82×1.33×45 min608K$1.38
51.92×1.3×46.2 min685K$1.57
62×1.25×48 min762K$1.76
72.06×1.2×50.1 min839K$1.96
82.11×1.14×52.5 min916K$2.15
92.14×1.09×55 min993K$2.35
102.17×1.04×57.6 min1.07M$2.54
112.2×1×60.3 min1.15M$2.73
122.22×0.95×63 min1.22M$2.93
132.24×0.91×65.8 min1.3M$3.12
142.26×0.88×68.6 min1.38M$3.32
152.27×0.84×71.4 min1.46M$3.51
162.29×0.81×74.3 min1.53M$3.70
172.3×0.78×77.1 min1.61M$3.90
182.31×0.75×80 min1.69M$4.09
192.32×0.72×82.9 min1.76M$4.29
202.33×0.7×85.8 min1.84M$4.48
212.33×0.68×88.7 min1.92M$4.67
222.34×0.65×91.6 min1.99M$4.87
232.35×0.63×94.6 min2.07M$5.06
242.35×0.62×97.5 min2.15M$5.26
252.36×0.6×100.4 min2.23M$5.45
262.36×0.58×103.4 min2.3M$5.64
272.37×0.56×106.3 min2.38M$5.84
282.37×0.55×109.3 min2.46M$6.03
292.38×0.53×112.2 min2.53M$6.23
302.38×0.52×115.2 min2.61M$6.42
312.38×0.51×118.2 min2.69M$6.61
322.39×0.5×121.1 min2.76M$6.81

Your task

The part that can’t run in parallel.

Reading and reconciling one worker’s report.

Shorthand such as 50k or 1.5M works.

Assumptions

System prompt and tools, before a worker does anything.

Split evenly across workers.

The coordinator reads every one.

Dollars per million tokens, from the site’s shared price table. Caching is ignored.

Don’t make every item wait

When work runs in stages, how items move between them matters as much as how many workers you have. Waiting for every item to finish one stage before starting the next means each stage takes as long as its slowest item. Passing each item along lets the slow ones overlap. Only wait when the next stage really needs all of the previous results, such as deduplicating findings.

Between stages

Finishes in 20 minutes. Every item waits at minute 10 for the slow first item, then the slow last item runs alone.

  • Find (stage 1)
  • Fix (stage 2)
Four items, two stages, 26 minutes of agent work. Item 1 takes 1 minute to find and 10 to fix; item 4 takes 10 to find and 1 to fix; the others take 1 minute each. Minutes along the bottom.

When not to delegate

  • The task is smaller than the handoff: writing the brief and reading the report costs more.
  • Every worker needs the same context, and it’s still changing.
  • The interface the workers share isn’t settled yet.
  • Integration eats the parallel savings, which the chart above shows as the downturn.
  • There’s no independent test to tell you which results to accept.

How this is calculated

Time. The best speedup n workers can get is Amdahl’s law: 1 / (s + (1 − s)/n), where s is the serial share. With 40% serial work and 4 workers that’s about 1.82×, and no number of workers passes 1 ÷ 0.4 = 2.5×. That’s the dotted line. The coordinator then reads and reconciles every report one after another, so the solid line adds the integration minutes per worker times the number of workers. That term grows while the parallel saving shrinks, which is why the curve turns down.

Tokens. One session reads the shared context and the unique work once. Fanned out, every worker pays its spawn overhead and reads the shared context for itself, the unique work is split between them, and the coordinator reads every report. Both write the same output.

Cost. Input tokens times the input price plus output tokens times the output price, from the site’s shared price table, added up before dividing by a million. Prompt caching would lower both sides and isn’t modelled.

Assumed. Work splits evenly, and workers never wait on each other or on a concurrency limit. The default minutes and token counts are illustrations; use your own.

Faster output still has to be reviewed: see how much agent output you can actually review.