Abstract
AI agents acting on a user's behalf — sending email, posting chat replies, creating tasks, calling tools, delegating to sub-agents — must decide, for each action, whether to act autonomously, escalate for a quick check, or hold for full human review. The prevailing mechanism is a single static confidence threshold ("auto-send if confidence > 0.8"), hand-tuned by operators. A single threshold is brittle: it cannot distinguish a low-stakes chat reply from a high-stakes external email; it does not adapt as the agent demonstrates competence (or begins to fail); and it forces every action type through one gate.
This publication discloses a method in which each action type maintains its own confidence band — an ordered pair (low, high) — rather than a single threshold. On completion of each action, the system records an outcome and recomputes that band, without human intervention, as a function of a sliding window of recent outcomes for that action type: the band tightens (raises low/high) as failures rise and loosens (lowers them) as successes accumulate. A subsequent action is routed by a trichotomy: confidence above high → auto-execute; below low → require human review; in between → escalate (a lightweight, time-boxed check). Because every action type calibrates independently, an agent may become fully autonomous on chat replies while still requiring review on external email — a graduated autonomy gradient that emerges from observed outcomes, not from manual tuning. The disclosure provides a system architecture, the detailed recalibration and routing mechanics, a data model, an enabling clean-room reference implementation, a fully worked example, a security and failure-mode analysis, standards/framework alignment, an evaluation methodology, and enumerated novelty claims (one independent plus fourteen dependent). It is published to establish dated, public prior art so the technique remains freely practiceable.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Dynamic Confidence Bands per Action Type", Technical Disclosure Commons, (June 29, 2026)
https://www.tdcommons.org/dpubs_series/10581