Hermes has introduced a new "bot mode" feature, enabling specialist AI agents to communicate and collaborate on shared tasks. This functionality allows for the creation of teams of specialized bots, each with defined roles, capabilities, and memory. This approach aims to enhance workflow efficiency by distributing tasks among agents that excel in specific areas, while also optimizing resource usage by removing unnecessary capabilities.

The Anatomy of a Bot

In Hermes, each bot is essentially a specialized profile. A bot's configuration includes its identity, brain, capabilities, memory system, and operations.

• Identity: This encompasses the bot's profile picture, name, title, and a description. The description is crucial as it informs the bot's persona and understanding of its role within a team. Hermes can automatically generate a soul.md file based on the provided name, title, and description, defining the bot's behavior and capabilities. This soul.md file can be further customized to establish team protocols, specify task delegation, and define limitations.
• Brain: This refers to the underlying model the bot uses. Users can select from various cloud-based models (e.g., GPT-4.5, Astra, Claude 3) or configure local models based on their device's specifications. This allows for flexibility in choosing the appropriate processing power for different tasks, potentially reducing costs by using less powerful local models for routine operations and more advanced cloud models for complex coordination.
• Capabilities: These are defined by skills and tools. Skills represent reusable workflows, while tools enable agents to perform specific actions (e.g., writing files, browsing the web). It is recommended to limit each bot's capabilities to those strictly necessary for its designated role to prevent unnecessary tool calls and improve efficiency. For instance, a researcher bot might need web browsing and file reading capabilities, while an orchestrator bot might not require direct web access.
• Memory System: This governs how bots retain information. Bots can have persistent built-in memory, storing information in memory.md and user.md files. External memory systems, such as Hindsight, offer more robust capabilities for cross-project insights but may not be necessary for all specialized bots. It is advisable to designate a single bot, like an orchestrator, to manage external memory to avoid conflicts.
• Operations: This section defines how bots execute tasks. This includes configuring the execution backend (e.g., Docker for isolated environments) and defining the workspace, which is the directory where bots can access and share files.

Building a Specialist Team

The bot mode allows for the creation of distinct specialist agents. In an experimental setup, three bots were configured:

1. Librarian: Responsible for organizing approved knowledge and managing an LLM wiki. This bot is equipped with skills for citation and wiki review.
2. Researcher: Tasked with investigating current reality questions, producing cited briefings, and identifying uncertainties or gaps in information. This bot is configured to avoid writing to the wiki directly and to save its research in a designated workspace folder.
3. Orchestrator: Acts as the central coordinator, clarifying outcomes, delegating research to the researcher, reviewing returned work for relevance and completeness, and managing revisions. The orchestrator is designed to delegate tasks and manage external memory.

Communication and Collaboration

Hermes bot mode supports several communication methods:

• Individual Bot Chat: Direct messaging with a single bot. These chats are permanent and can be compacted using the /new command.
• Bot-to-Bot Messaging: Agents can communicate directly with each other, allowing for task delegation and status updates. However, this can lead to infinite loops if not managed carefully.
• Group Rooms: These facilitate a shared conversation among multiple bots. Group rooms have a default turn limit to prevent endless loops and require human intervention to proceed. They allow bots to pool their tools and skills, creating a broader collective capability.
• Kanban: This interface offers more granular control over complex tasks, including ownership, revisions, and approvals, enabling the creation of intricate workflows.

Experimentation and Findings

An experiment was conducted to test the efficacy of a multi-agent team for a research task. The objective was to understand the state-of-the-art 2026 best practices for agentic workflows with multi-agent teams.

The orchestrator was prompted to initiate the research, delegating the task to the researcher. The researcher produced an initial report, which the orchestrator reviewed. The orchestrator identified a deficiency in the report, specifically the lack of a 2026 study, and requested a revision from the researcher. This iterative process, involving a human in the loop for approval, ensured the quality of the output.

Upon acceptance, the report was passed to the librarian, who then proposed adding it to the LLM wiki. The librarian's proposal was approved, and the information was integrated into the wiki, creating a new note on "multi-agent workflows" with citations and connections to related concepts.

Key takeaways from the experiment include:

• Specialization is Key: Limiting the capabilities of individual bots to their specific roles significantly improves efficiency.
• Orchestration is Crucial: A well-defined orchestrator can effectively manage team tasks, quality control, and revisions.
• Human Oversight: The ability to intervene and guide the process, especially in complex or iterative tasks, remains vital.
• Model Capabilities: While advanced models can perform complex tasks independently, multi-agent systems can leverage less powerful, potentially local models for cost-effectiveness, with a more intelligent orchestrator overseeing the process.
• Iterative Refinement: Bot mode requires ongoing refinement of instructions, tool selection, and memory configurations to optimize performance.

The experiment demonstrated the potential of Hermes bot mode for creating specialized agent teams that can collaborate to achieve complex objectives, with the flexibility to integrate human oversight and optimize resource utilization.