From one agent to an orchestrated team.
The same goal — a fast, consistent, defensible specials decision — re-architected as the constraints became clear. Step through each stage: the agents, the tools and sources in each pipeline, when they’re called, and the fallbacks built in. Stages 2.0 and 2.1 run the same real bid end to end. Stage 6 is the Product Portal path — the contract, the return trip, and the telemetry. Stage 7 is production today: the same engine on Claude Managed Agents, live since August 22, 2026, with the Copilot Studio build kept in step as the Azure digital twin.
Follow one bid through the engine.
A Knoll Joe Arm Chair special, from the dealer’s request to a priced decision in Product Portal.
One real request, followed stage by stage: enrichment and precedent, the governing rule, complexity, a sourced list price, and the decision landing in Product Portal. Plays when it scrolls into view and loops.
Step-by-step version ↗Open full screen ↗In production — AutoBid on Claude Managed Agents
The Claude build began in July as a hedge: a second implementation of the same engine, graded on the same verified bids, to keep the team from depending on one vendor’s roadmap. It turned out to be the stronger option. The cost comparison came out about 30% in Claude’s favor, and running two builds side by side caught four defects in the production logic before launch. So Claude took production. It went live on Saturday, August 22 at noon through Power Automate flows that carry the same Product Portal contract, and the Copilot Studio build became the Azure digital twin. In its first 30 days it evaluated 5,700 real dealer requests and agreed with the quoters’ actual decision 91.7% of the time. Select any component to see what it does.
Hover or tap an agent above to see its pipeline and what it returns for this bid.
◆Coordinator— Claude Sonnet 5 · sole decision-maker
One agent on Claude Managed Agents runs the proven five-phase logic in a fixed order: extract, enrich and find precedent, classify each deviation, score complexity, then price and format. It calls the pricing and knowledge-base tools itself and makes every decision itself. Its instructions are the canonical ones from the Copilot Studio build, carried over word for word, which is why a disagreement between the two builds points to a real defect. Since September its decisions use the Product Portal’s own vocabulary — Approve, Approve Standard, Approve Vary Easy, Declined (with or without an alternative), and Human Review Required — so accuracy is scored directly against what the quoters recorded.
- KnowledgeCanonical instructions + five phases · every requestcarried over verbatim from the Copilot Studio build
- SourceSpecials knowledge base — pages and attachments · per deviationRovo MCP, with a Confluence REST fallback when the MCP server is down; reads the pricing tables that live in attached files
- ToolDynamic Pricing API · once per orderpre-checked: a reference that can’t be priced is caught before the call
Why it matters · The pricing table was in a PDF attached to the article, not on the page. Once the engine could read attachments, the missing upcharge was found and the asterisk cleared. The page text always outranks an older screenshot on the same page.
