Rakuten engineer let Claude Code work alone for 7 hours on a vLLM feature

A Rakuten engineer asked Claude Code to add an activation vector extraction method to vLLM, a large open-source library. The agent worked for seven hours in one run with only occasional guidance.

Organization: Rakuten
Industry: E-commerce & Tech
When: 2025
Evidence: Vendor case study
Agents and tools: Claude Code

Reported outcome

Seven hours of mostly autonomous coding in one run. Rakuten reports 99.9% numerical accuracy versus the reference method.

7 hours
Autonomous run
99.9%
Numerical accuracy

Caveats

Published by the company that sells the agent. Treat numbers as marketing until confirmed elsewhere.

Published as an Anthropic customer story. No independent verification.
The engineer gave occasional guidance, so this was not fully hands-off.

What the agent does

Claude Code reads the vLLM codebase, plans the change, writes the code, and runs checks against a reference implementation. The engineer watched and gave short steering notes.

Why this works

The task had a clear, numeric definition of done: the output must match the reference method. A clear check lets an agent work for a long time without drifting.

What to copy

Before a long run, write the "done" test first. A numeric comparison is ideal.
Ask the agent to commit at checkpoints, so you can roll back a bad path.
Check in every hour or so. Steer early instead of fixing late.

Starter prompt

Written by us, not by Rakuten.

Goal: implement {feature} in this repository.

Definition of done: `python scripts/compare_reference.py` reports
agreement of at least 99.9% with the reference outputs in tests/reference/.

Work plan:
1. Read the relevant modules and write a short plan in PLAN.md.
2. Implement in small steps. Commit after each step that passes tests.
3. Run the comparison script after each major change.
4. If you are stuck for more than 3 attempts on the same error, stop
   and write what you tried in PLAN.md.