Optional defaultGMIInactivityOptional defaultOptional defaultGMIBaseOptional personaStrict validation enforcement configuration (optional, defaults to permissive).
Optional cognitiveOptional per-GMI cognitive memory factory used by devtools and advanced runtimes.
Optional shutdownHow long the manager waits for one GMI to shut down, in milliseconds, when
it shuts down, deactivates a session or cleans up idle GMIs. A GMI that has
not finished by then is logged at warn and finishes on its own while the
manager goes on. shutdown() shuts every GMI down at the same time, so it
takes about one bound however many GMIs are active. Defaults to 8000: inside
the 10 seconds docker stop waits before it kills a container, and above
the 5 seconds a GMI gives its metaprompt work to finish.
Optional learningThe runtime's learning layer, when the host turned learning on (AgentOSConfig.learning). With it, each GMI
the manager builds gets a learner session and the registry persona's traits, and sends the personality
section on every turn. Without it the manager builds GMIs as it did before the layer existed.
Optional agencyThe prompt-engine config of the GMIs createAgencySeat builds for the teams a GMI launches (launch_agency on the
full runtime). Their engine gets no utility AI, so it truncates history and makes no model call. AgentOS sets it
from its own promptEngineConfig when agencies is set.
Configuration options for the GMIManager.