Agentic Loops
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Agentic loops are an architectural pattern in which a large language model is embedded in a recurring execution cycle with access to tools, enabling it to take sequences of actions rather than responding in a single pass. Pedro Franceschi argues they are the defining structure of every effective AI product.
Core Argument: Agentic Loops as the Foundation of Good AI Products
Franceschi makes a strong positive claim about the universality of this pattern: "every single good AI product you've used is an agent loop with tools". From this perspective, the loop-plus-tools combination is not one implementation choice among many but the essential substrate of value-delivering AI applications.
Criticism: Over-Constraining the LLM Undermines the Pattern
A recurring critique in the surrounding discourse is that developers misapply or undercut agentic loops by treating the LLM as a "very precious thing that's very expensive" — wrapping it in excessive if-statements and tight context controls rather than allowing the loop to operate with sufficient freedom. This over-constraint is presented as the wrong approach that limits the value agentic loops can deliver.
Known Pitfall: Opaque Training Data Distribution
A structural limitation flagged for builders designing agentic loops is that LLMs have opaque training data distributions: "the biggest pitfall of LLMs is you have no sense of how much training data the model has seen for the exact thing that you're asking it". Because an agentic loop may route queries across a wide variety of domains and edge cases, this blind spot compounds — builders cannot reliably predict where the model will perform well versus poorly within the loop's operation.
Points of Disagreement
The material presents a tension rather than a direct debate between named speakers: the strong advocacy for agentic loops as the correct paradigm sits alongside two distinct criticisms — one behavioral (over-constraining the LLM) and one epistemic (unknown training data coverage). Both criticisms accept the loop paradigm implicitly but warn that naïve implementations will fail, suggesting the debate is less about whether to use agentic loops and more about how to deploy them responsibly.