Multi-Agent Requirements Engineering Assistant

context & problem

Requirements gathered from stakeholder interviews are often unstructured, contradictory, or ambiguous. Turning them into clear, testable requirements is a manual, expertise-heavy step that doesn't scale well with a single generalist prompt.

my role

Designer & builder — I designed the agent roles, their handoff protocol, and the validation logic that catches ambiguity before it reaches a human reviewer.

key decisions

1) Multiple specialized agents instead of one large prompt

Why? A single prompt asked to elicit, structure, and validate requirements at once tends to blur those steps together and misses contradictions.

Trade-off: More orchestration complexity and latency — vs — clearer separation of concerns and traceable reasoning at each step.

Impact: Each stage's output can be inspected and corrected independently, closer to how a real requirements engineering process works.

2) An explicit validation agent as a final gate

Why? Catching ambiguous or conflicting requirements automatically, before they reach a stakeholder review, saves the most expensive kind of rework.

solution highlights

• Elicitation agent that extracts discrete requirement candidates from raw notes/transcripts
• Structuring agent that formats them against a consistent requirements schema
• Validation agent that flags ambiguity, missing acceptance criteria, and conflicts

reflection

What worked well: splitting elicitation from structuring made both steps noticeably more reliable. What I'd improve: tighter feedback loops so the validation agent's findings can trigger a targeted re-elicitation instead of a full re-run.