Agent Task Break Detector

Agent Task Break Detector Synthetic training data for a small classifier that answers one question on every user turn of an AI agent session: Does this new message need the conversation so far, or is it a new task that a fresh session could handle just as well? Agent sessions build up context. When a user drops an unrelated task into a long session, the old history only gets in the way, so the clean move is a fresh session: /clear or /new in Claude Code, a new thread… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Agent-Task-Break-Detector.

About This Project

Agent Task Break Detector is a Hugging Face Space that serves as an interactive explainer and demonstration for a ModernBERT-based classifier published as danielrosehill/Agent-Task-Break-Detector-ModernBERT. The classifier answers a practical question in agent workflows: given a new message, does it belong to the session so far, or does it warrant a fresh context? The Space runs the int8 ONNX build of the model on CPU, using the input format and decision threshold of 0.76 specified in the model's new_task_split.json file.

The project is built with Gradio 5.49.1 and is distributed under the Apache 2.0 license. The underlying model is associated with a training dataset, danielrosehill/Agent-Task-Break-Detector, and the Space is tagged for agents, context management, topic shift detection, and ONNX inference. Source code is available within the Agent-Context-Harness repository on GitHub.

The Space may be of interest to developers building conversational or agentic systems who need to manage context windows, reset or prune session history when topics change, or evaluate lightweight topic-shift classifiers that can run without GPU hardware.

Project Details

Tags

task_categories:text-classificationlanguage:enlicense:cc-by-4.0size_categories:1k<n<10kformat:jsonmodality:textlibrary:datasetslibrary:pandaslibrary:polarslibrary:mlcroissantregion:usagentsai-agentscontext-managementtopic-shiftdialoguesynthetic

Explore More Projects