We reproduced a public ARC demo that checks, before any run, whether an agent's output file is inside its permitted write scope

What this check examines

Each role's permitted write scope and each task's output file (output_file) are written in separate settings. We checked whether this spike can find a mismatch between them before a run, by reading only the declarations in the config files. This is a reproduction of a public demo, not a check that started from a particular user's difficulty.

The design we checked

What we confirmed about the design is that it never starts CrewAI: it reads only the declarations in the config files and returns a mismatch with a reason code, so it can be tried without installing or launching an execution platform. We have not confirmed what was weighed when this design was chosen.

What we tried

We ran the demo at the public ARC commit 97a009bd (node examples/crewai-adapter-spike/demo.mjs). The permitted write scope declaration stayed unchanged; only the task output file (output_file) moved from a file inside the scope, to a file outside it, and back to the original file inside it. (1) Passed inside the scope, (2) failed with TASK_WRITE_SCOPE_OUTSIDE_AUTHORITY outside it, (3) passed again once restored. The exact file names used as input are in the README and the demo at the fixed commit.

The intended counter-example and what we restored

The failure in (2) is an intended counter-example: the output file was outside the permitted scope, so the check stopped it. We did not change the permitted scope declaration; only the output file went inside the scope, outside it, and back. It passed once restored. We did not look for failures this check would miss.

What was confirmed, and limits

What was confirmed is only that the declared permitted scope and the output file agree. On 2026-11-03 we reproduced the same fixed commit in a freshly fetched copy: the demo ran pass, intended failure, pass, and the test output reported 9 passed and 0 failed (Node v24.22.1). The test command exit code was not recorded. We have not confirmed that this holds for other versions, other configurations, a real CrewAI run, or different working conditions. CrewAI was not started and no tool was invoked or output file written, so whether a write outside the scope is stopped at runtime is not known.

What is unconfirmed in real use

This check covers setups that declare per-role write scopes and per-task output files separately. It does not cover setups with no write-scope declaration, or cases that need writes stopped while the agent runs. We have not confirmed how people who actually use such a setup experience it, how often it is used, or whether it led to any improvement.

What we want to check next

What we want to check next is whether declarations and behavior agree in a real CrewAI run. We have not tried it.

Limits

• This is a declaration consistency check, not a runtime safety control.
• Time spent and cost were not recorded and are unknown.
• Reproduced in a freshly fetched copy for one demo and one configuration at the fixed commit. Generalization to other versions, other configurations, a real CrewAI run or different working conditions, and any user experience, frequency or improvement, is unconfirmed.
• The test output reported 9 passed and 0 failed, but the command exit code was not recorded.