Agent Logic Is the Missing Layer in Enterprise AI

IBM Research shows that software primitives—not bigger models—are what make agentic AI actually work at scale.

By The ai-compute Desk

Bigger context windows are not the answer. IBM Research's new work argues that enterprise AI fails not because the underlying models are too weak, but because agents lack the structural guidance to navigate real workflows—and the token costs that pile up along the way are making pilots impossible to justify. [1]

The fix IBM calls 'agent logic' is a layer of software primitives—knowledge graphs, algorithms, program analysis libraries—that sit inside the agent harness and steer the model toward the relevant slice of a workflow, shrinking context before the LLM ever sees it. [1] Applied to mainframe code understanding across systems of up to 1M lines of code and 1K programs, the approach maintains marginally superior performance with ~30× lower token consumption than a baseline frontier LLM-only approach. [1]

The same logic applied to IT compliance automation—using adaptive planning and dynamic decomposition across specialized agents—boosts success rates from single digits to as high as +80% (Claude 4 Sonnet), while delivering 1.3–2.0× greater performance over fixed planning strategies. [1]

The pattern across every domain is the same: constrain what the model sees, and its outputs become cheaper, faster, and more trustworthy. Structural scaffolding does the work that prompt engineering never could.

Enterprise AI's bottleneck was never the model. It was the absence of anything telling the model where to look. [1]

Agent logic is software primitives, such as knowledge graphs, algorithms, program analysis libraries, which operate at the agentic layer (within an agent harness) and can intentionally steer the LLM in the direction of the enterprise workflow, reducing the context space.IBM Research