2025-09-14 –, Track 2
Rube is a universal MCP server that connects large language models to 500+ apps and manages massive context seamlessly. But the real story isn’t just what Rube does — it’s what we learned while building it. This talk dives into the engineering challenges behind production-ready multi-agent systems: managing large contexts efficiently, designing reliable tool schemas, routing user requests across hundreds of integrations, adding memory without bloating state, optimizing execution speed, and getting LLMs to “call the right tool at the right time.”
Through successes, failures, and plenty of debugging scars, you’ll learn practical strategies for building intelligent, scalable systems that push beyond toy demos. If you’re experimenting with MCP, LLM apps, or multi-agent orchestration, these lessons from the trenches will save you weeks of trial-and-error — and maybe inspire a new direction for your own projects.
Detailed Outline
1. Introduction
- Quick overview: what Rube is (universal MCP server, 500+ apps, large context).
- What the audience will gain: practical lessons, not just a demo.
2. The Problem Space
- Why we needed something like Rube.
- Challenges of connecting LLMs to many apps.
- Constraints: scale, reliability, speed, developer experience.
3. Lessons from Building Rube
Organized into key themes.
- Context Management
- Handling large prompts + responses without breaking limits.
-
Strategies: chunking, summarization.
-
Routing & Planning
- Teaching the LLM to select the right tool among 500+.
-
Designing tool descriptions that actually guide model behavior.
-
Execution & Optimization
- Making tool calls fast: caching, parallelization, error handling.
-
Integrating code execution safely and efficiently.
-
Memory & State
- Adding memory to Rube: session state vs. long-term memory.
-
What worked, what backfired, and how to keep memory useful but lean.
-
Developer Experience
- Debugging agent interactions (tracing, evals).
- Building confidence that the system will “do the right thing.”
4. What We’d Do Differently
- Missteps: over-engineering, underestimated bottlenecks.
- Hard-earned insights that could save others time.
6. Q&A / Discussion
OpenAI, Agents, Python
Target Audience –Intermediate
Jayesh loves exploring new technologies and putting them to work to solve real-world problems. He adapts very quickly to emerging frameworks or languages and plans his days to maintain a balance between sports, literary works, and coding. Jayesh enjoys speaking in public and interacting with people with different experiences than his own.