Code as Data: Why LLMs Fail at Structural Programming Tasks

Code is data. Why do AI coding agents pretend it isn’t?

George Ciobanu observed AI agents spending up to 40 seconds performing redundant grep searches during codebase exploration. This inefficiency stems from models treating structured logic as a linear stream of characters rather than a directed graph.

Why This Matters

AI agents currently operate on code as an unstructured stream of text, ignoring the explicit syntax trees and namespaces that compilers have utilized for decades. This text-first approach leads to hallucinated refactors where renamed functions leave behind broken callers, as the model lacks the structural root system required to track every edge in the code graph, resulting in production failures and 3am bug reports.

Key Insights

• Code functions as a graph where symbols, types, and calls connect across file boundaries through explicit syntax.
• Text-first LLM refactoring often misses indirect references, such as variables updated everywhere except inside a callback two levels deep.
• The pandō structural engine (2025) provides AI agents with a layer for namespaces, symbols, and dependencies to enable exact transformations.
• Traditional IDEs use tree-walking for ‘Rename Symbol’ to ensure unrelated identifiers remain untouched, a capability missing in standard LLM prompts.
• The Pando model represents codebases as single organisms where text surfaces are connected by a shared underlying root system.

Practical Applications

• Use Case: Integrating a structural engine with agents like Claude Code or Aider to find every caller of a method without hallucinations. Pitfall: Relying on grep-based search which may find matching strings that are not actual functional references.
• Use Case: Performing project-wide symbol renaming by updating nodes in a graph rather than strings in a file. Pitfall: Using find-and-replace on common names like ‘Alice’, which results in renaming unrelated variables in different scopes.

References:

• https://dev.to/ionut_georgianciobanu_08/code-is-data-why-do-ai-coding-agents-pretend-it-isnt-oj9