From Tool Calling to Loop Engineering: A Survey of Control, Feedback, Acceptance, and Self-Improvement in LLM-Based Agents

Abstract

Agents built on large language models (LLMs) combine a model with tools, memory, verifiers, and software that decides when work starts, stops, and changes. This survey traces how these responsibilities moved from the model's response into the surrounding harness between 2020 and 2026, and asks which component accepts a result and which components may change across runs. It synthesizes 171 sources and codes a purposive sample of 39 systems with a nine-field schema that separates checks available after deployment from oracles that exist only in evaluation. It also audits ten recursive self-improvement studies against a 12-item reporting checklist, and 18 further systems from an arXiv search. In 15 of 29 coded task-execution systems, the acting model ends the run by declaring completion, and a check that remains available after deployment takes part in acceptance in 11. Of the ten improvement studies, three state that the evaluator is fixed, one checks candidates for regressions on development data that the proposer does not see, none describes a rollback procedure, and six describe enforced sandboxes, limits, oversight, or access control, but only one restricts access to the evaluator. The spot audit found the same gaps in regression checks on unseen data and rollback. Controlled comparisons support gains from action-loop structure and interfaces, informative repair feedback, and evolving memory. Reliability and self-improvement claims should name the acceptance authority, the mutable component, the fixed outer controls, and the budgets. The schema and checklist provide a common format. Files in this record: - The paper (PDF).- Online Resource 1 (XLSX): the search and screening logs, the coding of the 39 systems, the checklist audit of the ten studies, the spot audit of 18 further systems, and the coding of the related surveys, with the supporting passage and locator for each code.