A registry of weighted, cooldown-gated, availability-gated curiosity actions, selected by weighted-random sampling only when an agent's primary work source is empty — producing intrinsic-motivation behavior without interfering with reactive operation.
This repository is a Technical Defensive Publication. It is published to establish dated, citable prior art for the techniques described herein.
- Publisher / Copyright holder: Gus IT LLC (Florida, USA)
- Author: Gustavo Assuncao, PhD
- Publication date: 2026-06-25
- Version: 1.0
- Document type: Technical Defensive Publication (public prior art)
- Classification: Public
- License: AGPL-3.0-or-later (copyleft; commercial license available)
By publishing this disclosure openly and with a date, we intend to place the described mechanism into the public domain of prior art, so that the technique remains freely practiceable by everyone and cannot be patented by others.
Autonomous, reactive AI agents — chat assistants, cognitive personas, and tool-using agents driven by a perceive–reason–act loop — sit idle whenever no message, task, or event is pending. That idle time wastes provisioned compute (GPU/CPU and a held model context) and, more importantly, forfeits opportunities for the agent to do useful proactive work: revisiting stale goals, summarizing accumulated context, learning a new skill, or surfacing a forgotten follow-up. Naively filling idle time with background activity is dangerous: undisciplined background tasks compete with reactive work, repeat themselves, and fire when their preconditions are not met.
This publication discloses an Idle-Triggered Curiosity Engine: a small, deterministic control component that activates only when the agent's primary work source reports empty, and then selects one intrinsic-motivation ("curiosity") action from a registry of action descriptors. Each descriptor carries three independent gates — a weight (relative selection probability), a cooldown (minimum time before the same action may re-fire), and an availability predicate (a runtime precondition function). Selection is a weighted-random draw restricted to the eligible set — descriptors whose availability predicate evaluates true and whose cooldown has elapsed. The chosen action is enqueued into the agent's ordinary task queue, so it flows through the same execution, safety, and budget machinery as reactive work and is automatically pre-empted the moment real work arrives.
The combination — idle-as-the-sole-trigger, plus a three-gate weighted registry whose selection is restricted to the runtime-eligible set — yields proactive, varied, non-repetitive, precondition-respecting agent behavior. This document fully describes the architecture, mechanics, data model, an enabling clean-room reference implementation, a worked example, failure modes, and the novelty claims.
We are publishing this as a defensive publication rather than seeking a patent.
The internal patentability screen for this technique rated it marginal (US 55/100, DE/EPO 42/100): the constituent ideas — weighted-random scheduling, cooldown timers, precondition gating, and reinforcement-learning "curiosity" reward — are individually well-trodden in game AI (utility AI, GOAP) and RL literature. Its narrow novelty (the idle-detection-as-sole-trigger combined with a three-gate registry restricted to the eligible set) is real but does not justify the cost and exclusivity of a patent.
Rather than leave the technique in a grey zone where a third party might later attempt to patent a close variant and assert it against practitioners, we choose to publish it openly with a date. This establishes prior art that:
- keeps the technique freely practiceable by anyone, forever;
- bars others from patenting this mechanism or trivial variations of it; and
- seeds a future open-source reference application (see docs/OPEN-SOURCE-APP.md) so the community can build on it.
This is part of a broader portfolio of defensive publications by Gus IT LLC covering agent-orchestration techniques.
| Document | What it covers |
|---|---|
| DEFENSIVE-PUBLICATION.md | The full whitepaper: architecture, mechanics, data model, enablement, worked example, failure modes, framework mapping, evaluation, and the complete novelty claims (1 independent + 14 dependent). |
| docs/PRIOR-ART.md | Fuller prior-art landscape, the delta table, and an honesty attestation. |
| docs/FIGURES.md | All Mermaid figures with captions (Figure 1–6). |
| docs/OPEN-SOURCE-APP.md | The planned open-source reference app and a generic Kubernetes/AKS deployment sketch. |
| src/ | Clean-room illustrative reference implementation (Node.js) demonstrating the novel mechanism, with a runnable self-check. |
| LICENSE / NOTICE | AGPL-3.0-or-later license and attribution. |
| CITATION.cff | Citation metadata. |
This repository is the technical seed for a planned, minimal open-source reference application — a sidecar/library that any reactive agent can adopt to gain disciplined idle-time proactivity. See docs/OPEN-SOURCE-APP.md for the design and a generic deployment sketch.
Licensed under the GNU Affero General Public License v3.0 (see LICENSE). The AGPL-3.0-or-later license includes an express patent grant, which is appropriate for an open prior-art release.
To cite this work, see CITATION.cff. Suggested form:
Assuncao, Gustavo. Idle-Triggered Curiosity Engine for AI Agents (Technical Defensive Publication, Version 1.0). Gus IT LLC, 2026-06-25.
Deposit: to be assigned (IP.com / Zenodo / arXiv) — establishes a public, dated, citable prior-art record.
Copyright 2026 Gus IT LLC (Florida, USA). Licensed under AGPL-3.0-or-later.