Agentic AI Solution Design and Multi-Agent Interaction
Domain 1 - Agentic AI solution design and multi-agent interaction. The exam tests whether you can pick an agent topology, define tool and message contracts, and run an interaction loop that terminates reliably instead of looping or drifting.
Design from the loop outward
An agent is a model plus instructions, tools, memory, and a control loop. Specify the loop before the prompt: plan, act, observe, reflect, stop. Every loop needs a hard stop - maximum turns, token budget, wall-clock deadline - and an explicit success predicate. Treat tool schemas as API contracts: name, description, JSON schema, error model, idempotency, and side-effect class. Structured output and argument validation before execution prevent a model from inventing parameters.
Choosing a topology
A single agent with good tools usually beats a swarm. Decompose only when subgoals need different context, permissions, or model sizes. Common patterns are supervisor/orchestrator-worker (one router, specialist workers, no worker-to-worker chatter), sequential pipeline, and reflection or debate with capped rounds. Give each subgoal exactly one owner. Inter-agent messages should be typed and minimal - pass artifacts by reference rather than replaying full transcripts.
Worked example
A refund request. A supervisor classifies intent and calls order-lookup and policy-check agents in parallel. Policy-check returns eligibility, reason, and citation. Only when eligible and the amount is under threshold does the supervisor call refund-executor; otherwise it escalates to a human queue. Reasoning: decision and execution are separated, so a misread policy cannot move money; the executor schema requires an eligibility token, making premature execution impossible; six turns and a token budget bound cost.
Traps
Over-decomposing into agents that only chat; no termination condition; letting the executor also decide policy; vague tool descriptions that cause argument hallucination; forwarding entire histories between agents; unbounded reflection loops; ignoring the token cost of multi-agent chatter.