Compound AI System
Quick Answer
A compound AI system is an automation system in which a foundation model is one component among many — retrievers, tools, memory, planners, verifiers, sandboxes, and approval gates — collaborating to complete a multi-step objective. The term, introduced by Berkeley AI Research in 2024, names the shift from single model calls to systems whose behavior emerges at the system boundary. It is broader than 'agent' and more specific than 'LLM application.'
Compound AI System
A compound AI system is an automation system in which a foundation model is one component among many — retrievers, tools, memory stores, planners, verifiers, schedulers, policy engines, sandboxes, and human approval gates — collaborating to complete a multi-step objective. The term was introduced by Berkeley AI Research in 2024 to name the shift from monolithic model calls to programs whose performance is produced at the system boundary rather than by any single model invocation. It is broader than "agent," which usually implies an autonomous loop, and more specific than "LLM application," which can be a single chat call. Characteristic properties include task decomposition, external grounding via retrieval, the ability to mutate state through actions, feedback loops with retries or replanning, and persistent memory.
See also
- Tool hijacking — failure mode specific to compound systems with tool brokers
- Indirect prompt injection — dominant attack class against systems that ingest untrusted retrieval
- Ambient authority — why models inside compound systems should not be principals
- Agent capability control — authority discipline that complements compound-system architecture