Project 7 — Multi-agent customer support

Requirements

Add specialists only for measurable policy/claim/document tasks, with clear coordinator ownership and escalation.

Architecture

Supervisor → scoped specialist subgraphs → typed reports → consistency check → response. Keep final writes outside specialist discretion.

Technologies

Existing graph/SDK, typed contracts, shared run budgets, scoped tools; reuse one infrastructure stack.

Database design

Parent/child run IDs, report evidence references, role/version tags, and conflict/escalation state.

API design

Public conversation/run API stays stable; specialist endpoints/contracts are internal and bounded.

AI design

Minimal per-role context, independent completion criteria, limited delegation depth, and explicit conflict handling.

Security

Distinct tool permissions; no automatic inheritance of broader supervisor credentials; controlled context sharing.

Evaluation

Compare against one agent on task success, support quality, cost, latency, and permission failures. Test contradictory specialists.

Observability

Delegation graph, role calls, shared budget, partial failures, and final ownership.

Deployment

Scale independent workers only when justified; preserve compatible report schemas.

Interview questions

• “Why multiple agents?” Answer: specialization must prove a benefit.
• “What if they disagree?” Answer: inspect evidence and apply authoritative rules or escalate; agreement alone is not truth.

Acceptance: Demonstrable benefit over the single-agent baseline with bounded coordination cost.