How to Build Planning Agents without losing control - Yogendra Miraje, Factset

9.5K views · Jul 23, 2025 · 15:57 min · Watch on YouTube ↗

Takeaway

Insert a natural-language blueprint step between user intent and planner to keep enterprise planning agents controllable, debuggable, and tractable.

Summary

FactSet distinguishes workflow-agent (static workflow run by agent) from agentic-workflow (dynamic workflow planned by agent) — keeping the spectrum between control and flexibility.
Adopts LLMCompiler architecture on LangGraph with nodes: blueprint generator, planner, executor, joiner; recursion limit prevents loops.
Key innovation: an intermediate natural-language 'blueprint' that limits the tool set fed to the planner, reducing context overload and cognitive load.
Aspect-based evals: LLM-as-judge for blueprint vs golden, code-based evals for tool selection, human-in-loop for report formatting; build MCP tool servers wrapping existing microservices.

planning-agentslanggraphmcp

Original description

LLMs are getting smarter—but Agents are still unpredictable, unreliable, and hard to control.

In this talk, I’ll share practical lessons from building real-world plan-and-execute agents —covering how to steer autonomous agents using agentic workflows, blueprints, and evals.

If you’re struggling to make your agents behave (without giving up flexibility), this one’s for you.

About Yogendra Miraje
I'm a backend engineer turned ML engineer turned AI engineer, with 16 years of experience building intelligent systems. I hold a Master’s degree in Computer Science from Northeastern University in Boston, and I currently work as an AI Engineer in FactSet.

I'm also the host of AI Blindspot, a podcast where we explore the frontiers of artificial intelligence—and the blind spots we often overlook in its development and deployment.

With a strong foundation in Machine Learning and software Engineering and a product-minded approach, I focus on aligning autonomous agents with real-world user goals, emphasizing safety, control, and robust evaluation techniques.

I'm passionate about building AI that’s not just powerful, but grounded, aligned, and truly useful in practice.

Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter