On a whiteboard, an agent is prompt + tools + model. In production, that equation grows a long tail: session store, streaming, state management, context compression, memory, file upload, human-in-the-loop, code execution, tracing, auth, stop/resume, multi-agent handoff, evals, security review. The first agent absorbs that cost quietly. The third one doesn’t — by then most teams have three conversation stores, three security reviews, three observability schemas, and no way for one agent to hand work to another.

This talk presents MicroAgents: a microservices-style architecture for agent fleets, drawn from the platform running ThoughtSpot’s AI analytics in production for enterprise customers. Each team owns exactly two things — an MCP server exposing its domain’s tools, and a prompt template encoding its expertise. Everything else belongs to a shared base agent service. Onboarding a new agent becomes a config entry rather than a codebase.

The session covers the microagent contract and why defining it early matters, two orchestration patterns (sub-agent as tool, and full handoff), treating memory as scoped organizational infrastructure rather than per-agent state, collaborative planning across sub-agents, and zero-instrumentation tracing. It also covers build-versus-buy decisions on managed runtimes, why framework choice matters less than the abstraction above it, and where MicroAgents diverge from the Agent Skills model.

Attendees will leave with a concrete architecture for making the marginal agent cost weeks instead of quarters.