The creator demonstrates building a robust multi-agent workflow using Hermes Agent integrated with a Kanban board, enabling autonomous coordination of specialized agents to research, decide, build, and produce deliverables with a single human approval checkpoint. This system ensures efficient task management, parallelism, self-healing, and full auditability, and has been open-sourced as a versatile framework for scalable multi-agent applications.
In this video, the creator demonstrates how they successfully built a robust multi-agent workflow using Hermes Agent combined with a Kanban board system. Multi-agent workflows are often challenging due to coordination issues, but this setup allows a fleet of agents to autonomously conduct research, make decisions, build tools, and produce deliverables with only a single human approval gate at the end. The Kanban board acts as a centralized coordination layer, preventing agents from duplicating work or conflicting with each other by managing tasks as cards that agents claim, work on, and then mark as done. This design ensures durability, parallelism, event-driven task flow, self-healing capabilities, and full auditability.
The Kanban board is the core innovation here, serving as the single source of truth for all agents. Each task is represented as a card with a title, assignee, and status, stored in a simple SQLite file. The dispatcher continuously claims ready tasks and spawns the corresponding agent to work on them in isolated environments. Tasks can depend on others, automatically waiting for parent tasks to complete before progressing, enabling parallel work without conflicts or the need for complex messaging systems. This architecture solves the common problem of agents stepping on each other’s toes in multi-agent systems.
The creator designed a specific workflow to identify real pain points faced by AI agent users by scouting complaints from various web sources, validating and scoring them against a rubric, and then routing them to either build a fix or create explanatory videos. The workflow involves multiple specialized agents such as scouts for research, an orchestrator for decision-making, researchers, analysts, builders, testers, and video producers. The orchestrator deduplicates issues, scores them, and decides the next steps, while the human gate in Telegram allows the creator to approve, shelve, or modify proposals before any building or video production begins, ensuring control over resource usage and quality.
A live demonstration shows the workflow in action, with scout agents gathering data, the orchestrator evaluating and routing tasks, and multiple researcher agents working in parallel to verify and contextualize issues. The system autonomously generates build proposals and video outlines, which are sent to the creator for approval via Telegram. Upon approval, builder and tester agents autonomously develop and validate tools or video content. The system also includes self-healing features that detect and fix issues like misplaced deliverables without human intervention, highlighting the robustness and automation level achieved.
Finally, the creator shares that the entire workflow template has been open-sourced for others to adapt to their own needs. This generalized framework supports research, scoring, decision-making, and execution phases with a human approval checkpoint, making it a versatile foundation for various multi-agent applications. The video concludes with an invitation for viewers to share their own multi-agent setups and workflows, emphasizing the ongoing challenges and opportunities in building durable, scalable agentic systems. The creator plans to continue exploring and improving multi-agent workflows in future content.