Hermes Bot Mode: I Built a Team of AI Agents That Hand Off Work to Each Other

What if your AI agents behaved less like isolated chatbots and more like a team of specialists that could actually collaborate? 🤖 That’s what I wanted to test with the newly released Hermes Bot Mode desktop plugin. Instead of manually switching between different Hermes profiles, copying context, and triggering every stage myself, I built a small team of three AI agents: Researcher → Risk Analyst → Thesis Editor Then I gave one task to one agent. The goal was simple: see whether the agents could gather evidence, challenge each other’s work, and produce a final answer through agent-to-agent handoffs without me manually operating every stage. And it mostly worked. But the interesting part isn’t the stock research demo itself. It’s what Bot Mode changes about how we interact with Hermes Agent, persistent AI agents, and multi-agent workflows. 🎥 Full video walkthrough 🧠 What Is Hermes Bot Mode? Hermes already supports profiles. A profile can have its own configuration, model settings, soul, memory, skills, and tools. That means you could already create multiple specialized Hermes agents. Bot Mode doesn’t replace that system. Instead, it adds a visual usability and orchestration layer on top of Hermes profiles. The easiest way I can describe it is: Hermes Profiles give you multiple isolated brains. Bot Mode gives those brains faces, rooms, and a team interface. Instead of remembering profile names or managing everything through the CLI, you get a visible roster of bots. Each bot can have its own: 🪪 Name and visual identity 🎯 Specialized role 🧠 Personality and memory 🛠️ Tools and skills 💬 Persistent conversation 🔄 Ability to communicate with other bots That last capability is where things get interesting. 🧪 The Experiment: Can 3 AI Agents Complete One Research Task? I wanted a task where simply generating an answer wasn’t enough. The workflow needed different types of judgment. So I created three specialized bots. 🔎 Bot 1: Stock Market Researcher The first agent gathers evidence. Its job is to research two U.S.-listed companies in the same sector using public filings and reputable market data. For the demo, I compared NVIDIA and AMD across areas such as: - Revenue growth - Margins - Balance-sheet risk - Valuation context - Business catalysts But I didn’t want the first agent’s answer to automatically become the final answer. That is where the second bot comes in. ⚠️ Bot 2: Risk Analyst The Risk Analyst is deliberately adversarial. Instead of expanding the researcher’s conclusions, its job is to attack them. It checks things like: - Are the reporting periods comparable? - Are the numbers supported? - Are valuation assumptions reasonable? - What downside risks are missing? - Is there contradictory evidence? - Does any language sound like an unjustified forecast? This creates an important separation between generating research and reviewing research. 📝 Bot 3: Thesis Editor Finally, the reviewed material goes to the Thesis Editor. Its job is not to research everything again. It takes the evidence and critique from the previous agents and produces a balanced comparison containing: 📚 Sources 📅 Relevant dates 🎚️ Confidence notes ❓ Known unknowns ⚖️ A more balanced final analysis Now we have a simple multi-agent pipeline: Research → Challenge → Synthesis 🔄 The Interesting Part: I Only Talked to One Agent This was the actual test. I gave my instructions to the Stock Market Researcher. That was my only initial task instruction. I wasn’t manually taking its output and pasting it into the Risk Analyst. I wasn’t opening the Thesis Editor and telling it what to do next. Instead, the researcher gathered its evidence and then attempted to hand the work to the Risk Analyst. The Risk Analyst reviewed it and attempted to pass the reviewed material further down the chain. The bots were composing and routing these handoffs themselves. That’s a much more interesting interaction model than: Human → Agent A → Human → Agent B → Human → Agent C Instead, we’re moving toward: Human → Agent A → Agent B → Agent C → Human The human defines the objective while specialized agents handle parts of the coordination. 👀 Persistent Conversations Make the Workflow Easier to Understand One thing I like about Bot Mode is that these aren’t just invisible background calls. Each specialist exists as a recognizable bot with its own persistent conversation. I can open the researcher and inspect what it did. I can open the Risk Analyst and see what it received. I can inspect what happened during a handoff. That matters when you’re experimenting with multi-agent AI systems. If an AI team produces a bad result, you don’t just want the final answer. You want to understand: Where did the workflow go wrong? Was the original research weak? Did the critic miss something? Did information disappear during a handoff? Did the final agent overstate the evidence? Making the agents and their conversations visible gives you a much better mental model of the system. 🛠️ Creating Specialized Bots Creating a bot from the desktop interface is fairly straightforward. You can give it a: - Name - Title - Description - Avatar The advanced configuration is where the specialization becomes more powerful. You can configure the bot’s soul, assign skills, control its available tools, or clone it from an existing Hermes profile. That means these don’t have to be three copies of the same generic assistant with different names. You can design genuinely different specialists. For example: Research Agent Web access + research skills + evidence-focused instructions. Critic Agent Verification instructions + skeptical personality + strict rules around unsupported claims. Editor Agent Strong synthesis instructions + limited mandate to introduce new claims. The architecture becomes interesting when the agents have different responsibilities, context, tools, and behavioral instructions rather than simply different labels. ⏰ Bots Can Also Run Recurring Jobs Another useful feature is scheduled jobs. You can configure a bot with an instruction and schedule it to run at a particular frequency. For example, the stock research bot could potentially run a recurring morning research task. That opens up use cases beyond manually initiated conversations: 📊 Daily market research 📰 News monitoring 🔍 Competitive intelligence 📈 Recurring business analysis 📋 Periodic reporting 🧭 Research updates Persistent specialists become much more useful when they can perform recurring work instead of waiting for a new chat every time. ⚡ But Bot Mode Is NOT a Full Workflow Engine This distinction is important. It’s easy to see agents communicating and assume you’ve suddenly built a full multi-agent orchestration system. That’s not what Bot Mode currently is. The handoffs are real, but they’re per invocation. A receiving bot might respond later, and a bot already processing something might not be interrupted immediately. Bot Mode also does not guarantee parallel execution. My demo was effectively sequential: Researcher → Risk Analyst → Thesis Editor So I wouldn’t treat Bot Mode as a replacement for a proper DAG or structured workflow engine. 🐛 It’s Also Still Beta I actually encountered a failed handoff during the demo. The Risk Analyst completed its review but failed to successfully pass the work to the Thesis Editor. Rather than hiding the failure, I kept it in the video because it demonstrates an important limitation. Bot Mode is currently beta release. Failures can happen. In my case, I instructed the researcher that the previous delegation had failed and asked it to retry. The researcher then communicated with the Thesis Editor, passed along the reviewed material, and the workflow continued. For production-grade autonomous workflows, that distinction matters. A good agent interface doesn’t automatically give you guarantees around: - retries - dependencies - execution state - parallelism - deterministic routing - failure recovery Those are orchestration problems. 🆚 Hermes Bot Mode vs Hermes Profiles vs Hermes Kanban These three concepts solve different problems. Hermes Profiles provide the underlying specialization. They give agents separate configurations, model settings, memories, souls, tools, and skills. Hermes Bot Mode makes those specialists easier to operate as a visible team. You get identities, persistent conversations, a roster, and agent-to-agent communication. Hermes Kanban is still more appropriate when I need a structured project with explicit tasks, dependencies, and organized multi-agent collaboration. I have used Kanban extensively for more complex AI-agent workflows because those tasks need more structure. So I wouldn’t think about Bot Mode as: “Bot Mode replaces Hermes Profiles or Kanban.” I would think about it as another interaction model. Profiles = specialization 🧠 Bot Mode = persistent specialists + communication 💬 Kanban = structured task orchestration 🗂️ The right choice depends on the workflow. 💡 Why I Think This Direction Is Interesting Most AI assistants still revolve around one interface: one user ↔ one chatbot But many real tasks aren’t naturally one-role problems. Consider startup research. You might want: Market Researcher → Competitor Analyst → Skeptic → Founder Memo Editor For software development: Architect → Developer → Reviewer → QA Agent For content: Researcher → Scriptwriter → Fact Checker → Editor For sales: Lead Researcher → Account Analyst → Outreach Writer The value isn’t simply having “more agents.” Adding ten agents to a workflow doesn’t automatically make it better. The interesting question is whether we can give specialists clear responsibilities and useful handoffs while keeping the system understandable to the human operating it. Bot Mode is an interesting step toward making that experience more accessible. I’m particularly interested in where this model goes next. Because the bigger opportunity isn’t just giving AI agents better answers. It’s giving us better ways to organize,