Skip to content
What Agents Pick

do AI coding agents build their own instead of using a tool?

mostly yes. in 219 of 355 scored runs (62%) Claude Code added no third party tool for the job and wrote it itself. highest: auth 35 of 35, jobs 33 of 33, caching 36 of 36, storage 34 of 36. lowest: payments, 0 of 35.

from 355 runs of Claude Code across 3 models and 10 categories. the prompt never names a tool. updated Oct 18, 2026.

get the new results by email

one email a week at most. new agent test runs, one chart, and which tools agents picked. free. read issue 1

built its own, by category

all models pooled. a run counts when no third party package or service was added for the job.

  1. auth35 of 35100%
  2. jobs33 of 33100%
  3. caching36 of 36100%
  4. storage34 of 3694%
  5. email24 of 3667%
  6. deployment21 of 3658%
  7. database20 of 3656%
  8. monitoring11 of 3631%
  9. ORM5 of 3614%
  10. payments0 of 350%

built its own, by model

  1. Opus 5.571 of 12059%
  2. Sonnet 5.571 of 11860%
  3. Haiku 5.577 of 11766%

other questions

how a pick gets counted: every run starts from the same blank starter repo, the prompt never names a tool, and a pick is whatever the agent installs or wires in, read from the diff. full rules on the methodology page, raw data in runs.json.