As AI agents become more capable, a new challenge is emerging in enterprise environments: understanding how agents arrive at their decisions. Enterprises interact with large volumes of complex data, this makes transparency in decision-making pathways especially important and the lack thereof detrimental to system improvements.
Today’s agents do much more than answer questions; they query databases, retrieve documents, call APIs, and combine information from multiple sources to complete complex tasks. This process, often called data planning, determines how an agent sequences these operations to satisfy a user’s request.
The problem is that planning is often hidden from the user’s view. An agent receives a prompt, performs a series of actions, and returns an answer. When the result is incorrect, slow, or incomplete, it can be difficult for a business owner to understand what went wrong and how to better guide the agent. Organizations need agents whose reasoning processes can be inspected, evaluated, and improved.
This lack of visibility creates a major obstacle for enterprise adoption. Teams need to debug failures, compare planning strategies, optimize performance, and evaluate reliability. Without insight into how an agent executes a task, these activities become significantly harder.
L.A.K.E.: An Agentic Data Planning Framework
At Megagon Labs, we explored this challenge through L.A.K.E. (Logic Agent for Knowledge Extraction), an agentic data planning framework designed to make planning observable.
L.A.K.E. converts natural language questions into executable workflows that operate across heterogeneous data sources. Rather than relying on a single planning strategy, it supports multiple planning regimes and exposes their execution through interactive Directed Acyclic Graphs (DAGs) with step-level provenance.
By exposing execution paths, intermediate outputs, and tool dependencies, L.A.K.E. enables teams to evaluate not only whether an answer is correct, but also how it was produced.
What We Learned About AI Agent Planning
Planning strategy matters.
Single-shot tree planning achieved the strongest overall performance and lowest latency, while iterative planning offered stronger opportunities for self-correction at a significantly higher execution cost. We also found that planning structure itself can meaningfully influence agent behavior.
These findings suggest that planning is not simply a reasoning problem. It is a systems design decision that affects efficiency, reliability, and operational cost.
Towards Explainable and Observable AI Agents
As agentic systems become more deeply integrated into enterprise workflows, observability will likely become as important as capability. The future of enterprise AI may depend not only on building smarter agents, but also on building systems that can be understood, evaluated, and improved.
L.A.K.E. reflects this broader direction: building AI systems that make their planning visible, helping developers understand not only what an agent did, but why it did it.
📄 Read the Full Research Paper
Find L.A.K.E. on Github
Written by Jean-Flavien Bussotti and Megagon Labs
