~/talk

Building distributed multi-agent systems

on YouTube 01h 54m
Talkat Google Cloud Next '26 · 2026-04-22..04-24· [ai-assisted-development, serverless]

speakers

name:Priya Pandey
links:none on file
name:Shir Meir Lador
links:none on file

resources

video:https://youtu.be/VjBijrS19gY
slides:none on file

summary

Walks through building a self‑correcting Course Creation Pipeline that runs each specialist as its own Cloud Run service, using the Agent Development Kit and the A2A protocol to discover agents. Researcher agent calls Google Search, a Pydantic-powered Judge returns deterministic JSON with pass/fail, a LoopAgent iterates until max iterations: 3, then a SequentialAgent hands approved research to a Content Builder. Local end‑to‑end test passes using Gemini 3 Flash preview; Cloud Run deploys are shown but require correct IAM and storage roles.

AI-written summary · about

chapters (57)

  1. 0:00Introduction to building multi-agent systems
  2. 1:18Meet the speakers and session overview
  3. 2:47Why multi-agent systems matter in production
  4. 6:28Introducing the AI course creator demo
  5. 8:20Architecture of a distributed multi-agent system
  6. 9:06Key technologies for scalable agent systems
  7. 10:51Live demo of the AI course creator in action
  8. 11:35The researcher agent and its role
  9. 12:40The judge agent and quality gates
  10. 14:50Debugging agents with the ADK web UI
  11. 15:31Loop agent architecture for iterative research
  12. 16:14The Content Builder agent and its workflow
  13. 17:19How agents communicate via Agent-to-Agent protocol
  14. 18:20Understanding agent cards for service discovery
  15. 20:28Managing state and context with the orchestrator
  16. 21:11Setting up the lab environment in Cloud Shell
  17. 23:04Cloning the starter repository and project setup
  18. 25:22Enabling required Google Cloud APIs
  19. 28:20Installing dependencies with UV and pip
  20. 30:31Configuring environment variables for the lab
  21. 31:34Why UV is faster than pip for dependency management
  22. 34:15Defining the researcher agent in code
  23. 36:44How the researcher agent uses the Google Search tool
  24. 38:56Prompt engineering for agent behavior control
  25. 39:16Choosing the right model for agent tasks
  26. 40:23Defining the judge agent with structured output
  27. 44:18Understanding judge agent autonomy and constraints
  28. 45:46Agent autonomy as a dial not a switch
  29. 46:28Testing the judge agent with research findings
  30. 47:12Audience topic selection for agent testing
  31. 48:57Evaluating the researcher agent's output
  32. 51:04Defining the Content Builder agent and its instructions
  33. 55:24How the Content Builder uses shared session state
  34. 57:12Persistent vs short-term memory for agents
  35. 57:57Defining the orchestrator agent and its callbacks
  36. 1:00:04How the orchestrator manages session state
  37. 1:02:54Connecting agents using the Agent-to-Agent protocol
  38. 1:04:25Defining remote agents in the orchestrator
  39. 1:05:49Agent authentication and IAM in Cloud Run
  40. 1:07:39How the escalation checker controls the research loop
  41. 1:10:04Understanding the escalation checker's role
  42. 1:12:15Creating the research loop with researcher and judge
  43. 1:12:58Building the sequential agent pipeline
  44. 1:14:46Testing the local agent pipeline before deployment
  45. 1:16:51Deploying agents to Cloud Run for production
  46. 1:21:05Live testing of the deployed multi-agent system
  47. 1:21:26Monitoring and observability for agent systems
  48. 1:24:20Troubleshooting Cloud Run deployment issues
  49. 1:34:04Answering audience questions about multi-agent systems
  50. 1:39:28Debugging approaches for agent workflows
  51. 1:43:21Managing memory and context limits in agents
  52. 1:45:32State sharing and intelligent decision-making in agents
  53. 1:46:44Stress testing agents before deployment
  54. 1:48:11Scaling agents with Cloud Run's auto-scaling features
  55. 1:51:10Troubleshooting build and deployment permissions
  56. 1:51:34Recap and next steps for building multi-agent systems
  57. 1:53:00Next episode preview and closing remarks

related talks

~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.