Author(s): Dr Swarneendu AI

Originally published on Towards AI.

The Reason Isn’t the Model.

I debugged a production agentic recently.

The rest of the article explains why enterprise agentic AI projects fail in production: the true cost scales quadratically with the number of turns because each additional step re-sends more context and accumulates loop overhead. It walks through the difference between single-pass and agentic loop cost functions, highlights common production breakers—turn count drift, context pollution, and context-window “roofline” crashes—and shows that these issues often don’t appear in staging. Finally, it outlines practical fixes that optimize the loop (stateless turns with explicit state injection, hierarchical planning with single-shot execution, and hard caps with graceful degradation) and argues that teams need a “harness” cost model to avoid systematic approval based on misleading staging estimates, ultimately improving the odds of reaching production.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI