The Limits of Conversational AI

The past few years have seen large language models become genuinely useful conversational partners. But pure conversation has a fundamental ceiling: a model can tell you how to book a flight, yet it can’t actually book one.

Agentic AI is the move from talking to doing.

Defining Agentic AI

An agentic AI system is one that can:

  1. Perceive its environment — ingest information from the web, APIs, sensors, file systems
  2. Form a plan — decompose a high-level goal into an executable sequence of sub-tasks
  3. Take action — invoke tools (search engines, code interpreters, external APIs)
  4. Sustain execution over a long horizon — persist through multi-step tasks without hand-holding

The critical distinction from conversational AI is the execution loop: perceive → plan → act → observe → re-plan, rather than simply respond.

Core Building Blocks

Planning

Modern agent systems typically use an LLM as the “reasoning core,” paired with prompting strategies like Chain-of-Thought or ReAct (Reasoning + Acting) to decompose complex tasks.

Tool Use

Agents interact with the external world through structured tool interfaces — function calling APIs, shell execution, database queries, and any number of service integrations. A capable toolset is what makes the difference between a clever planner and something that can actually change the state of the world.

Memory

  • Working memory: the current execution context — usually the LLM’s context window
  • Long-term memory: vector databases or structured stores for retrieval across sessions
  • Episodic memory: logs of past task runs that enable reflection and self-improvement

Multi-Agent Coordination

A single agent hits capacity limits on complex tasks. Multi-agent frameworks (AutoGen, CrewAI, LangGraph, and others) allow specialized agents to divide labor: an orchestrator, executors, and critics each handling what they do best.

Open Research Questions

[TODO] Fill in the lab’s specific research directions in Agentic AI here.

Safety and reliability remain the central open challenges in this space. How do you keep autonomous agents within well-defined boundaries? How do they degrade gracefully when something goes wrong? These are questions the lab considers worth sustained investment.