A platform group must ship three internal agents in one quarter: a log triage agent, a documentation agent and a release-notes agent. All three need the same core behaviour, namely an agentic loop that calls tools, handles errors and manages a growing context. The group has two engineers and no appetite for maintaining shared infrastructure of its own. Which construction approach best fits that staffing constraint?
- AWrite one bespoke harness in-house that all three agents import, implementing the loop, tool dispatch and context handling to the group's own specification.
- BBuild the log triage agent on the Claude Agent SDK and hand-roll the other two, since documentation and release notes are simple enough to run as single model calls.
- CRun the three agents as subagents beneath a coordinator agent that decides which of them handles each incoming request and supervises the work they return.
- DBuild all three agents on the Claude Agent SDK, which supplies the loop, tool invocation and context handling, and write only the tools and prompts that differ between the three. Correct
Why A is wrong: Sharing one harness across three agents is sound instinct, but the group then owns and maintains that loop indefinitely, which is precisely the shared infrastructure burden two engineers said they cannot carry.
Why B is wrong: Some tasks genuinely are single-call workflows rather than agents, but the stem states all three need a tool-calling loop with error and context handling, so two hand-rolled loops are being written for no stated reason.
Why C is wrong: Coordination is a real pattern, but the three agents serve separate purposes and no routing problem is stated, so this adds a supervisory layer the task has not earned and still leaves the loop unbuilt.
Why D is correct: The SDK already implements the behaviour common to all three agents, so the two engineers write only the differentiating tools and prompts and inherit maintenance of the loop rather than owning it.