A Flexible Multi-Agent LLM-Human Framework for Fast Human Validated Tool Building

TL;DR

The paper introduces CollabToolBuilder, a four-agent (Coach, Coder, Critic, Capitalizer) LLM–HITL framework that iteratively develops reusable tools for complex scientific document generation. It combines Reinforced Dynamic Prompting, a problem-environment memory, and a semantic tool library to balance autonomous tool-building with targeted human guidance, validated through a 30-document experimental corpus. Key contributions include open-source datasets and tooling (DEA, HumanLLM), and evidence that hybrid HITL-then-auto configurations yield faster convergence and higher-quality outputs. The work demonstrates domain adaptability and sets the stage for broader applications in iterative problem solving beyond scientific writing.

Abstract

We introduce CollabToolBuilder, a flexible multiagent LLM framework with expert-in-the-loop (HITL) guidance that iteratively learns to create tools for a target goal, aligning with human intent and process, while minimizing time for task/domain adaptation effort and human feedback capture. The architecture generates and validates tools via four specialized agents (Coach, Coder, Critic, Capitalizer) using a reinforced dynamic prompt and systematic human feedback integration to reinforce each agent's role toward goals and constraints. This work is best viewed as a system-level integration and methodology combining multi-agent in-context learning, HITL controls, and reusable tool capitalization for complex iterative problems such as scientific document generation. We illustrate it with preliminary experiments (e.g., generating state-of-the-art research papers or patents given an abstract) and discuss its applicability to other iterative problem-solving.

Figures (3)

• Figure 1: General tool building workflow and common internal workflow of each human guided LLM agent
• Figure 2: Coach step interface (1st iteration).
• Figure 3: Capitalized function example (2nd iteration).