Tim Handley

Agent swarming lets autonomous AI sub-graphs collaborate mid-execution, sharing intermediate findings, triggering follow-up investigations and building on each other's work in real time. This session share how this can work in production. A single agent with tool calling answers a question by reaching for sources one at a time. Fan that out into a supervisor dispatching agents in parallel, and you cover multiple sources at once — but the agents still work in isolation. Swarming changes that, granting sub-graphs the ability to talk to one another while they run. The session covers how sub-graphs share intermediate state to spin up cross-graph investigations concurrently, how clarification-aware states pauses pending, and the patterns that keep swarming predictable: time-budgeted execution, call limits, iteration ceilings and full-stack observability ‚— and even a role for a supervisor. Drawing on a 22-agent agentic analytics platform helping over 600 schools securely solve complex, multi-source data problems to improve student outcomes, this session shows how to make agent swarming powerful, bounded and trustworthy.