Multi-Agent Systems at Scale as a Shared Platform for the enterprises

An agent is simple: Prompt + Tools + Model + Boilerplate. The first three are where product teams create value. The last one - state management, history compression, streaming, cancellation, tracing, memory, persistence - is 80% of the code but 0% of the differentiation.
At ThoughtSpot, we built an Agent Platform that draws a hard line between agent logic and agent infrastructure, letting product teams ship customer-facing agents faster by owning only what matters: their prompts and their tools.
This talk covers the infrastructure patterns behind that separation:
State management across tool calls. Stateless tools (state on the agent, passed as arguments) give you testability and let the LLM reason about state. Stateful tools (state in the tool service) avoid serialization overhead. I'll walk through flow diagrams, show how we propagate state via tool response metadata, and discuss when each pattern fits.
Configuration-driven agent definitions. Agents defined entirely through config - templated prompts, tool endpoints, sub-agent rules, compression strategies. Teams ship agents without writing orchestration code.
Inter-agent communication. Two patterns: agents-as-tools (sub-agent called like any tool, returns structured output) and agent handoff (full conversation transfer). The platform handles routing and context - teams just declare delegation rules.
Shared memory across agents. Memory in the platform, not individual agents, means knowledge accumulates across agent boundaries. Tiered scoping (tenant, org, user) with retrieval that surfaces relevant context regardless of which agent captured it.
Tool protocol design. MCP as the base, with patterns layered on top: cancellation semantics, progress streaming, context variable propagation, and adapters for existing services.
Building for customer-facing scale adds constraints - high concurrency, encryption, tenant isolation, auditability - that shaped our API design throughout.

Speaker: Ashish Shubham, ThoughtSpot

Attend in person or online.  See weblink