F4 - agents & tool use
AGENT
│
┌──────┼───────┐
▼ ▼ ▼
Search Model Ticket API
- an agent can potentially invoke an authorized tool that actually performs the action
User request
↓
Agent
↓
decides tool required
↓
create_ticket(...)
↓
tool executes
↓
result returned
↓
Agent responds
tool
- a capability exposed to the agent that performs the action when invoked by the model
- eg:
get_detector_status(detector_id)
search_manual(query)
get_experiment_schedule(date)
create_maintenance_ticket(detector_id, reason)
-
current external state/dynamic state can be provided by a tool
-
eg: current detector temperature
-
agents can perform multi-step tasks
-
they must still be given least privilege permissions
-
humans can be in the loop, eg: manual human approval
-
tool failures should be handled appropriated:
- retry if transient
- report unavailable
