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

Introduction and Definitions 00:03

• The presenter, Yogi from Factset, introduces the topic of building planning agents and the challenge of balancing control with autonomy in AI applications
• Highlights recent exponential growth in AI but notes enterprise agents often lack proper context, especially for enterprise-specific workflows
• Defines key concepts: an augmented LLM (LLM + tool + memory), workflow (static, controllable), agent (autonomous, flexible), agentic workflow (dynamic, agent-driven)
• Clarifies differences between a workflow agent (static workflow, agent executes) and agentic workflow (agent plans and runs, workflow is dynamic)

Agentic Workflows: Importance and Architecture 03:07

• Agentic workflows are critical for automating enterprise processes at scale while leveraging existing microservices investments
• Emphasizes the agentic spectrum: agentic workflows have more "agenticness" than workflow agents
• General applicability of the concepts beyond just enterprise context

Building Planning Agents: Concepts and Patterns 04:07

• Trend has been reliance on react-based agents; for agentic workflows, a switch to proactive agents is needed
• Core components: tools, memory, reflection, and especially subgoal division (task decomposition)
• Highlights references to research papers and suggests a Langchain blog and code as resources
• Presents a practical architecture (LLM compiler): user input goes to a blueprint generator (high-level plan), planner (low-level tasks), executor (executes tasks), and joiner (combines outputs)
• Introduces re-planning logic to control execution loops and response delivery

Tooling and Microservices Integration 06:18

• Components such as blueprint generator, planner, executor, and joiner can be structured as nodes (e.g., in LangGraph)
• Integration with enterprise microservices is crucial; design tools tailored to those services
• There is no one-to-one mapping between tools and microservices; design should be flexible
• Important to "think like the agent": equip the agent with thorough knowledge of tools and microservices
• Recommends using standards like MCP for tool servers and to provide agents with tool purpose, description, input/output contracts, and validation checks for safety

Blueprint Generator and Planning 08:02

• Blueprint is a step-by-step natural language workflow, guiding the planner and reducing cognitive load
• Blueprint offers improved control over task planning, limits tools exposed to the planner, and helps manage context window limits
• Aids interpretability of agentic behavior and facilitates collaboration with non-technical users

Example Workflow and Evaluation 09:37

• Demonstrates an agentic workflow for preparing for a company's earnings call (e.g., NVIDIA)
• Steps: summarize previous earnings call, retrieve current financial data, suggest questions, and generate a comprehensive report—each step has linked tools/functions
• Shows improvement in response structure and usefulness after implementing agentic workflow
• Stresses the need for rigorous evaluation (evals): maintain both component and end-to-end evals, use various strategies (code-based, LLM as judge, human-in-the-loop), and focus on metrics important to your application (e.g., blueprint correctness, tool selection)

Limitations and Good Practices 11:55

• Agentic workflows are not suitable for all scenarios: fixed/repetitive tasks (use ETL), workflows difficult to capture, need for deterministic outcomes (compliance/safety), or low-latency/cost environments
• Best practices: start simple and build complexity gradually, use blueprints to structure tasks and limit tool exposure, design user-friendly tools, implement safety guardrails and observability, and adhere to sound software engineering principles

Key Takeaways and Q&A 13:13

• Agentic workflows are planned and executed by agents, bringing scalable reliability
• Subgoal division is a crucial design pattern; "plan and execute" is the central agentic architecture
• Build tools to complement and leverage existing microservices; adapt architectures as needed
• Experiment with research approaches and prioritize evaluation frameworks
• In response to questions: Langchain and related GitHub projects are good starting points for agentic workflows, and orchestration frameworks (MCP, LangGraph, etc.) should be chosen based on context and organizational needs—there is no one-size-fits-all solution