Data Science Dojo

Episode 5: Multi-Agent Systems & Workflow Orchestration with @SambaNovaSystems

The webinar, part of a series on AI agents, focused on building multi‑agent systems and orchestrating workflows with SambaNova’s platform. Lead AI architect Quasi Ankama from Canva walked the audience through design patterns, tooling, and practical implementation. Ankama explained why single‑agent pipelines quickly hit context limits, accumulate errors, and suffer from diluted expertise. He introduced three core multi‑agent patterns—supervisor/worker coordination, parallel fan‑out, and writer‑critic loops—that address these issues by enabling specialization, concurrency, and built‑in quality checks. The session used an incident‑postmortem use case, where a supervisor agent spawns log‑analysis, metric‑analysis, and reporting sub‑agents. A live demo showed how the Deep Agents library generates isolated task tools from a simple dictionary, allowing each sub‑agent to run with its own model and prompt while returning concise summaries to the supervisor. By adopting these patterns, enterprises can scale AI‑driven processes, keep token usage low by matching models to tasks, and maintain auditability through isolated contexts. SambaNova’s fast inference and tooling lower operational costs, making large‑scale multi‑agent deployments feasible for thousands of users.