Multi-Agent Architecture
Split agent work across specialized roles — planner, validator, executor, critic — so each component has a narrower job than a single monolithic agent.
A multi-agent system assigns distinct agent roles to stages of the loop. A minimal split is plan generation, plan validation, and plan execution under plan-validate-execute. Reflection setups often add a separate critic: one agent plans and acts, another evaluates after each step or batch (Reflexion).
Specialization helps because each role can use different prompts, models, or tool subsets from the shared tool inventory; see multi-agent system advantages for how this architecture drives modular upgradability, diverse perspectives, and parallel execution. Parallel candidate-plan generation with an evaluator is another multi-agent latency/cost tradeoff. Roles that only read or advise can multiply freely, but keep exactly one writer per piece of state (single-threaded writes); a shared task board is one way to sequence specialists when several must write. Treat role boundaries as harness design: clear inputs/outputs between agents matter as much as each agent's internal ReAct-style loop.
Roles and delegation frameworks (AutoGen, CrewAI, and similar) push the same idea further: assemble agents each with role, goal, backstory, and tools, then delegate toward an overall goal. Process shapes include sequential pipelines, hierarchical directors (a manager routes among specialists — AutoGen's group chat manager is this pattern), and consensual collaboration where agents jointly decide how work proceeds. A minimal unit is the Assistant–UserProxy pair. In software engineering domains, roles specialize into orchestrator, driver/programmer, reviewer/navigator, tester, and information retriever; see navigator-driver pair programming for decoupling implementation plans from code synthesis. When agents own durable assets rather than one-shot tasks, see stateful task agents; when the model chooses the next task rather than a fixed graph, see LLM-driven workflow routing.
Sources
- AI Engineering (Chip Huyen), ch. 6
- Prompt Engineering for LLMs (Berryman), ch. 9