Federated Agent Architecture: Leveraging Client Context & Cloud Inference
Agentic AI, AI Infrastructure, Context engineering, LLMOps, MCP
Drawing from our real-world experience building rtrvr.ai web agent platform, we present a contrarian architecture that moves agentic execution from the cloud to the client, while retaining inference and orchestration in the cloud.
We'll dive deep into:
- Architecting Federated Executions: How we leverage client-side resources (unblocked bandwidth, user credentials) for reliable execution while performing context engineering and AI inference server side.
- Sub Agent Context Coordination: Strategies for managing agent memory, context, and state across asynchronous, complex workflows that can run for 30+ minutes.
- Enabling the "Agentic Web": How this architecture unlocks cross-application workflows handled by a web agent client, ie: a Slackbot filing a Jira ticket via a web agent added with an MCP URL.
Building rtrvr.ai, the SOTA AI Web Agent that autonomously completes tasks, creates datasets from web research, and integrates any APIs/MCPs – with dead simple prompting and your own browser!
Pioneered vertical federated learning @ Google
M.S. Computational Engineering, M.S. Computer Science