[Feature Request] Add organization-level opt-out for system prompt injections blocking multi-agent workflows
Bug Description
Subject: Opt-out request — system-injected "Do not call the AgentTool unless the
user requested it" is degrading a documented multi-agent workflow
My org (e803cefb-fd8c-46f5-aefe-284b6184a1fd) appears to be enrolled in an experiment
that injects this into the system prompt on every request:
"Do not call the AgentTool unless the user requested it"
"Do not use workflows or deep-research unless the user requested it"
I am the account owner and I want to opt out. I could not find any supported way to do
so: no settings.json key, no environment variable, no managed-settings.json key, no
org-level control in the console, and no documentation that this is an experiment at
all. Please either opt my org out or tell me the supported mechanism.
WHY IT MATTERS — measured, not anecdotal:
My repo is built around a documented orchestrator that delegates recon and
implementation to subagents (8 project subagents in .claude/agents/, a routing ladder
in CLAUDE.md, artifact contracts, skeleton-based navigation). Since the injection
started, subagent spawns per session went 8.2 (Jun 1-Jul 15) -> 5.7 (Jul 16-31) ->
0.9 (August). A -89% drop.
The second-order effect is the real damage. With subagents, intermediate work (file
dumps, greps, exploration) collapses into a short summary in the main thread. Without
them, all of it lands in the main context. My sessions went to 46-48 MB with 4-6
compactions each. Every compaction discards the anchors already read, so the work gets
redone. The restriction does not just prevent delegation — it actively degrades output
quality by forcing context exhaustion.
This is not specific to me: another user publicly reported the identical directive
"superseding nearly all efforts to get subagents to work."
WHAT I WANT: opt my organization out of this experiment, or document a supported
per-user/per-org way to disable it. A blanket behavioral restriction with no opt-out
and no documentation is not a reasonable default for accounts that have deliberately
built multi-agent workflows.
Environment Info
- Platform: win32
- Terminal: null
- Version: 2.1.240
- Feedback ID: 9a5d5843-12e6-49c2-9d0d-90628b31eb97
Errors
[]