Autonomous Agents vs. the Rest of the Spectrum
Over 230,000 organizations have adopted Copilot Studio. In the last three months alone, they created more than one million agents, covering over 90% of the Fortune 500. But Smith draws a firm line between the spectrum of agents. A basic RAG lookup is not an autonomous agent. The distinction: autonomous agents respond to events (an email arrives, a database row changes, a schedule fires) without a human initiating the conversation. They reason, plan, and act. Building one requires stronger orchestration, observability, and governance from day one.
Ten Customer Lessons That Shaped the Roadmap
Cram 20 documents, 10 connectors, and an entire sales department into one agent and it fails. Smith’s fix: build specialist agents, one task each, then wire them together.
“So building these SME’s is key to this outcome, to reliability.” — Ray Smith
That insight drove multi-agent orchestration, released at Build. Nine more customer lessons produced nine more features: o1/o3 reasoning for contract and M&A analysis, MCP servers to wrap 40-parameter APIs, agent flows for deterministic sub-tasks, activity streams for debuggability, and centralized governance so IT can cap capacity and audit every run.
Suncorp’s RAG Deployment: Two Million Interactions, 50% Fewer Escalations
Suncorp fields two million customer contact center interactions every quarter. Before RAG, staff typed keywords into a search box, got a list of ten articles, and guessed which one answered the caller’s question (all while the customer waited on hold). Shen’s team replaced that with a vector search plus Azure OpenAI pipeline that returns a cited answer in seconds. Pilot results: 50% drop in escalations to next-level support. A separate single-view-of-claims tool has since processed over three billion input tokens and one billion output tokens.
Designing for Production: Governance Tradeoffs in Agentic Systems
Full autonomy is not the right answer for every step. Shen frames the core tradeoff: a deterministic flow gives you confidence every required step runs, but it hits walls when something unexpected arrives. A planning agent handles more edge cases but introduces variability. Most production systems need both: a master agent watching the end-to-end process, specialist sub-agents for defined tasks, and an explicit exception path back to a human or a dedicated exception-handling AI. Auditing is not optional. Regulators will ask which data drove each decision, and the answer needs to exist in logs.
Live Demo: Multi-Agent Claims Processing End to End
Smith ran a four-agent claims pipeline live. A Claims Manager Agent receives the trigger and calls three sub-agents in sequence: a Policy Researcher (o3 deep reasoning over the policy document and police report), an AU10TIX Validation Agent (license, registration, and repair shop ID checks via partner API), and a Decision Analyst. The Comms Agent then either emails an approval directly to the customer or fires a Teams adaptive card to a human adjuster. The auto claims manager live run (52:37) shows the full activity stream, drillable to each tool call and chain-of-thought prompt.
Notable Quotes
So garbage in is garbage out. Ray Smith · ▶ 12:41
So building these SME’s is key to this outcome, to reliability. Ray Smith · ▶ 14:13
this application has processed over three billion input tokens Jon Shen · ▶ 38:05
Key Takeaways
- Multi-agent systems beat single uber-agents: focused SME agents per task deliver reliable, debuggable outcomes.
- Email was 47% of agent triggers six months ago; custom API connections have cut that to 17% as maturity grows.
- Governance tooling (activity streams, analytics, capacity caps) is now the top ask from production-scale customers.