AI · AGENT ORCHESTRATION
A multi-agent eCommerce support system on Elasticsearch Agent Builder that resolves tickets through either a single autonomous agent or a six-agent pipeline with adversarial peer review.
↓ SCROLL
eCommerce support teams burn most of their time on the same resolvable tickets, and it is slow and expensive. In the README's baseline, a human handles a ticket in about 84 minutes at roughly $28, and coverage stops outside business hours. SupportGenius targets that common, repetitive slice: the issue categories a system can classify, research, and resolve deterministically, around the clock.
SupportGenius is an orchestration layer built on Elasticsearch Agent Builder, with seven specialized agents and ten custom tools. It runs in two modes. A streamlined autonomous agent resolves simple issues in a single API call. A six-agent orchestrated pipeline (Triage, Research, Decision, Simulation, Execution, Quality) handles harder tickets across 16+ calls, and runs an adversarial peer review before committing a response. Triage classifies complexity up front and sets an adaptive token budget, so effort scales with the ticket instead of being spent flat. In the README's figures this lands at about $0.45 per ticket versus the $28 human baseline, with subsecond processing and roughly 85% coverage of common issue categories.
The incoming ticket is classified by complexity. Simple tickets skip the simulation phases, moderate ones get two research phases, and complex ones get three plus trending analysis. The classification sets a per-agent token budget.
Research agents query Elasticsearch across support tickets, customer profiles, product catalog and knowledge base, using index searches and ES|QL queries to assemble context.
The Decision agent drafts a proposed resolution from the assembled context.
Two isolated agent conversations, an Optimist and a Pragmatist, debate the draft using separate conversation IDs to prevent prompt contamination. Convergence is detected automatically; disagreements resolve through weighted scoring that prioritizes VIP customer status.
The Execution agent applies the agreed action, synthesizing new tools at runtime when a novel action type is required.
The Quality agent checks the outcome, and the knowledge base is populated automatically so similar tickets resolve faster next time.
A single autonomous call for straightforward tickets, or a 16+ call orchestrated pipeline with adversarial review for complex ones.
Optimist and Pragmatist agents argue opposing positions in isolated conversations, exchanging quoted argument blocks, before a response ships.
Ticket complexity controls research depth and whether simulation runs, and per-agent token consumption is tracked and displayed.
Budget exhaustion falls back to previous outputs instead of failing the pipeline.
New tools are registered at runtime to handle action types the system has not seen before.
Resolved tickets populate the knowledge base so common issue categories get faster over time.