What this video shows
The first session in this ODSC stream is Nouha Dziri's keynote on teams of agents for software engineering. It asks whether a system can improve difficult engineering outcomes by assigning more work at inference time, rather than relying only on a larger model or one longer response.
Use the talk to form a testable design, then compare one agent, repeated samples and a coordinated team on the same task set. More agents can add search and critique, but they can also duplicate work, propagate errors and spend more tokens without improving accepted outcomes.
Read the official ODSC session description for the speaker, level and 25-minute session scope. Use Learnetto's AI agent test-set guide to compare agent configurations with stable cases and release thresholds.
What you will learn
- Compare agent teams with repeated single-agent sampling, not only with one baseline run.
- Count accepted task outcomes, elapsed time, tokens, tool calls and review effort.
- Give agents distinct roles only when the task has separable work or useful independent checks.
- Stop adding agents when coordination cost grows faster than accepted-task quality.
How to apply this safely
- Choose ten representative engineering tasks with known acceptance checks.
- Run one agent, repeated independent attempts and a coordinated team under the same tool and token limits.
- Record merge conflicts, duplicated work, incorrect consensus, elapsed time and human review.
- Keep the smallest configuration that meets the quality threshold, and preserve a fallback for tasks that need deeper search.
Important limitations
- The YouTube item is a full conference stream, and this guide covers its first 25-minute keynote rather than every later session.
- A conference presentation cannot establish that one agent topology works across repositories, models or tool environments.
- Extra inference-time compute raises cost and latency, and coordinated agents can share the same mistaken assumption.
- The official schedule describes the session, but teams still need task-level measurements from their own codebase.
Sources to check
- ODSC AI West 2026 Virtual Sessions, Day 1 Track 2 Official ODSC stream published October 27, 2026.
- Scaling Inference-Time Compute with Teams of Agents for Software Engineering Official session page for Nouha Dziri's 25-minute keynote.
- ODSC AI West 2026 schedule Official schedule confirming the October 27 session time, speaker and level.
Continue learning on Learnetto
Build an AI agent test set
Create stable tasks, transformations and release thresholds.
Backtest AI agents
Replay historical cases before changing production traffic.
Reduce coding agent cost
Measure cost per accepted task and route work by difficulty.