A generative AI help desk has an inventory specialist agent and a shipment specialist agent. A query can need one specialist or both.
Pass inputs such as the product ID and question only to the specialists needed, then receive their results. The same orchestrator agent must assemble the final user-facing answer every time.
Which arrangement best fits these responsibilities and result flow?
The agents-as-tools pattern lets an orchestrator invoke required specialists and combine their returned results. Compare it with parallel independent answers, an unconditional fixed sequence, and autonomous handoffs by tracing responsibilities, inputs, and return paths.
Detailed explanation
Autonomous handoffs are another coordination approach, but this arrangement does not retain the required final-answer owner.
Peer handoffs can suit tasks whose next specialist is chosen dynamically. In this particular arrangement the last specialist responds directly, so the required orchestrator no longer assembles every final answer.
This distributes work but does not satisfy selective specialist use or a single answer assembled by the orchestrator.
Independent work can be parallelized, but this arrangement runs inventory even for a shipment-only query and sends separate answers. It fails the stated requirements for selective specialist calls and centrally assembled output.
This keeps specialist selection and answer assembly with the orchestrator while giving specialists bounded responsibilities.
The Strands Agents agents-as-tools pattern allows an orchestrator to invoke a specialist and use its response. Passing the product ID and question to the required specialist and returning its result fits this scenario selection and answer-ownership requirements.
A fixed sequence can suit known dependencies, but it needlessly runs specialists that this query may not require.
A fixed sequence is useful when work follows known dependencies. Requiring every query to pass through both specialists performs unnecessary work here, so it does not match the selective-call requirement.
Check official documentation
Compare the following points with the official documentation.
Compare AIF-C01 Task 2.1 with the Strands Agents agents-as-tools, Graph, and Swarm concept documentation. Trace inventory-only, shipment-only, and combined invented queries on paper, listing specialists called, inputs, result recipients, and the final answer owner. Do not run an SDK or model.
Official references
- Official documentation 1
- Official documentation 2
- Official documentation 3
- Official documentation 4
- Official documentation 5
What to confirm
Explain selective specialist calls that return results to an orchestrator. Distinguish them from unconditional execution of all specialists, autonomous handoffs, and automatic history or memory sharing.
Key points
- Responsibilities
- Calls and returned results
- Answer assembly
- Information sharing
Notes
- Environment: Reference-based concept review of current AIF-C01 Task 2.1 using invented systems; no accounts, live integrations, or SDK execution.
- This tests roles and communication, not SDK parameters, default execution limits, or a claim that every environment uses the same shared-memory implementation.
- Use current Strands Agents concept documentation rather than instructions for new adoption of Bedrock Agents Classic, which is closed to new customers.
Foundation review
Separate specialist and orchestrator roles
With agents as tools, the orchestrator can invoke the specialists it needs and use their results in its answer. Explicit responsibilities, accepted inputs, and returned outputs help avoid duplicated work and unnecessary information sharing.
Distinguish dependencies from autonomous handoffs
A graph can define dependencies and branches; graphs are not all unconditional fixed sequences. Compare that with autonomous handoffs such as a Swarm and choose according to the required answer owner and coordination behavior.
Make information exchange explicit
Creating several agents does not automatically share every history or persistent memory. Specify inputs, returned results, and stopping conditions, and evaluate how added coordination affects quality, latency, and cost.