Subagents vs Agent Teams? 🧠 Hermes Bots, Goal Loops & Kanban Graphs
This video explores different levels of agent workflow complexity, from single agents with subagents to multi-agent teams using Kanban task management. The presenter tests when delegation, goal loops, and specialized agent teams make sense versus when a single capable agent is sufficient.
Summary
Callum (WanderLootz) examines the question of when to build multi-agent teams versus relying on a single agent with delegation capabilities. He introduces a progression of complexity: a single agent, a single agent with delegated subagents, agents with goal loops for quality checking, specialized agents in bot mode with persistent memory and skills, and finally Kanban-based task graphs for managing complex workflows.
The presenter demonstrates that delegating tasks to subagents can significantly reduce token consumption for the parent agent—from 102,000 tokens to 37,000 in his example—because subagents work in parallel and return compressed summaries rather than flooding the parent agent's context window. However, subagents start with no knowledge of the parent conversation history and only receive information from the task definition and context fields.
Goal loops are introduced as a way to improve single-agent output quality without adding complexity. By setting a clear goal with success criteria, an agent can iterate up to 20 times, checking if work meets standards and adjusting as needed. The presenter demonstrates adding criteria like requiring peer-reviewed sources and watching the agent autonomously refine its work.
Bot mode allows creating specialized profiles with persistent memory, specific skills, and tools. Unlike subagents that disappear after completing a task, bot profiles retain knowledge and can learn and improve over time. Group chats enable coordination between multiple specialized bots (orchestrator, researcher, librarian) through message exchanges, though limited to three exchanges before requiring human input.
Kanban is presented as a task graph system where the orchestrator doesn't block waiting for results but launches named profiles with persistent memory to work independently. Tasks persist in a SQLite database, recovery processes can retry failed tasks, and tasks can be locked requiring human intervention. The presenter walks through creating a Kanban board, enabling auto-decomposition where a triage specifier refines raw ideas and a decomposer breaks them into subtasks, and assigning work to performer profiles.
Testing different approaches on the same research task shows that single agent with delegation and orchestrator without auto-decomposer used similar tokens, while auto-decomposer with goal loops added overhead for simple tasks. The presenter emphasizes that the right approach depends on specific needs: start simple and add complexity only when additional coordination actually solves problems.
About this episode
Is one CAPABLE agent better than a TEAM of agents? It depends 👀 There are many tools, workflows & strategies depending on your specific needs. In today's video, I explore the different levels of agentic AI coordination using Hermes, from single agents with subagents to bot mode to full kanban task graphs. ✨ https://www.patreon.com/cw/wanderloots Agentic AI is getting better all of the time: the models are more powerful, the harnesses are more efficient, and the context management (knowledge in Obsidian & agentic memory) is enabling more fine-tuned customization for each of us. But knowing when to use which tools... that can feel a little overwhelming. Today, I walk through the core levels of agentic coordination: 1) a single agent vs agent with delegated subagents 2) running iterative loops (goal mode) 3) the new Hermes Bot mode (specialized agents) 4) the Kanban Plugin (aka task graphs) 5) and finally, testing how they all compare By the end of the video, you'll have a better understanding on when you should be doing what, along with the knowledge of how to structure your own complex workflows when the need calls for it. My hope is that this video helps you not only understand how to use bots, but also the bigger picture of what this means for working with agents across any workflow and any agent. For more on Hermes: Hermes Agent, by Nous Research (https://hermes-agent.nousresearch.com/), is a self-evolving AI Agent harness. That means it has skills, tools, and memories that enable the agent to tap into past conversations, codify style & preferences, & build custom workflows that improve every time you use them. Hermes has persistent memory, which means it extracts key information and remembers it over time, building a profile of you & how you work. These memories are used to build skills that improve over time as you repeat your workflows. Over time, you'll need to ask Hermes to do less, while still achieving more (without repeating yourself). For more on connecting to Ollama: https://www.youtube.com/watch?v=4KXLW9Y1r4c I also touch on why the skills & architecture of Hermes make Obsidian & the LLM Wiki a natural companion of this incredible agentic AI tool. More on LLM Wiki here: https://youtu.be/n4EVksU_EOs I hope you enjoy! ✨ P.S. I greatly appreciate any feedback, please let me know what you think 😊 Join My Membership: YouTube https://www.youtube.com/channel/UCFiU1vIpPD3lQltke_18m3A/join Patreon: https://www.patreon.com/c/wanderloots 💌 Sign up for my [free newsletter: Recalibrating](https://paragraph.xyz/@wanderloots.eth?referrer=wanderloots.eth) 🏡 Wander my Digital Garden https://wanderloots.xyz Timestamps: 00:00 Subagents vs Teams? 00:38 Today's Outline & Goals 01:08 Agents, Loops & Graphs 02:21 Enabling Subagents 03:11 Testing Subagent Delegation 05:41 Comparing Delegation & Non-Delegation 08:05 Goal Mode aka Loops 09:03 Testing A Goal Loop 11:59 Model Routing & Auxiliary Models 13:58 Reviewing The Goal Output 16:13 Bot Mode & Intro To Graphs 18:57 Examples To Help Visualize Graphs 22:27 Kanban vs Delegated Subagents 24:15 Enabling Hermes Kanban 25:48 Testing Hermes Kanban 26:39 Dispatcher, Specifier & Auto-Decomposer 28:47 Project Boards 29:40 Kanban Orchestration Settings 30:31 Kanban Setup: Research Example 32:20 Running The Kanban Board 36:19 Reviewing Kanban Results & Debrief 38:10 Orchestrator Decomposer (Manual, Not Auto) 41:05 Cost & Quality Comparison 42:04 What's Next? LINKS (MY WORLD) 🧭 [Recalibrating Newsletter Home](https://paragraph.xyz/@wanderloots.eth?referrer=wanderloots.eth) 🏔️To start reading from the beginning: [Recalibrating Newsletter Entry 1: What Recalibrating Means To Me](https://wanderloots.substack.com/p/1-what-recalibrating-means-to-me) 🌍 My [Website](https://wanderloots.com/) 📸 My [Print Shop](https://wanderloots.darkroom.com/) ✨ SOCIALS 🟣 [Farcaster](https://warpcast.com/wanderloots.eth) 📸 [Instagram](https://www.instagram.com/_wanderloots/) 📰 [Flipboard](https://flipboard.com/@_wanderloots) 📍 [Pinterest](https://www.pinterest.ca/wanderloots/) 🐦 [X (Twitter)](https://twitter.com/_wanderloots) 🤖 [Reddit](https://www.reddit.com/user/_wanderloots) MY FAVOURITE TOOLS 😴 🤯 The [Waking Up App](https://dynamic.wakingup.com/guestpass/SC4914439) (use this link for a 30 day free trial) 📝 [Obsidian](https://obsidian.md/) (decentralized note-taking) 📹 [Adobe Suite](https://prf.hn/l/lQ9DwpA) (general creativity) ❤️ [Welltory](https://app.welltory.com/payments/plans/main/?coupon=wanderloots) (Health Tracker) EQUIPMENT USED 6. Camera [Sony A7iii](https://amzn.to/3seSHv6) 7. Lens [Sony F2 28 mm](https://amzn.to/3TiWCT2) 8. Tripod [K&F Concept](https://amzn.to/3soCKCP) 9. Main Lighting Neewer 660 PRO RGB: https://amzn.to/3CEcU2V
Key Insights
- Delegating tasks to subagents reduces the parent agent's context consumption from 102,000 tokens to 37,000 because subagents work in parallel with compressed context windows and return synthesized reports rather than raw data
- Subagents have no knowledge of the parent agent's conversation history or prior context—they only receive information from the target and context fields populated by the parent when calling task delegation
- Goal loops allow a single agent to iterate up to 20 times checking if work meets defined success criteria, enabling quality improvement without multi-agent complexity by effectively linking multiple queries into a larger task
- Bot profiles in Hermes retain persistent memory, skills, and context across sessions unlike subagents, allowing agents to develop specialized expertise and learn from experience over time
- In Kanban task graphs, the orchestrator can launch tasks asynchronously without blocking, whereas in delegated task systems the parent process is blocked and cannot work until the subagent returns a result
- Kanban stores all task data in SQLite databases making everything permanently accessible for review and editing, unlike delegation which can lose context during compression
- The decomposer model for breaking down tasks into subtasks benefits most from high-intelligence models while execution requires less intelligence, making it cost-effective to pair a powerful planning model with weaker execution models
- Testing showed that for simple research tasks, a single agent with delegation and an orchestrator without auto-decomposer used similar token amounts, while auto-decomposer with goal loops added overhead
Topics
Transcript
[0:00] This is indeed a very effective way to significantly reduce costs. That's it. It just dropped from 45,000 tokens to 18,000. One agent can already delegate tasks to subagents. So what does creating a team of agents actually give you? This is the question I get asked most often about multi-agent workflows. Why build a team when one capable agent can handle the job? And, honestly , sometimes it's not worth doing. One powerful agent with subagents is often enough. But sometimes a more complex workflow can improve quality and efficiency. Hi, my name is Callum, also known as [0:30] WanderLootz, and welcome to today's video about agents, cycles, and graphs. When are multi-agent teams worth it? This process…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Wanderloots
How I Upgrade My Hermes Agent Team 🧠 Bot Mode, Skills & MCP Tips
Callum demonstrates how to upgrade a specialist researcher agent in Hermes by equipping it with custom skills, access to a Zotero reference library, and integration with Consensus for academic paper discovery. The upgraded researcher then collaborates with an orchestrator and librarian agent to produce high-quality research reports on agentic memory poisoning, showcasing how improving individual agents enhances team performance.
Always On Agentic AI? 🧠 Hermes Supercomputer Setup (Abacus AI)
Callum demonstrates how to deploy and run the Hermes AI agent 24/7 in the cloud using Abacus AI's Supercomputer, a managed cloud platform that eliminates the need for personal computer resources. The setup process is automated, requires just one click to install, and includes features like memory management, daily automations, and integration with messaging apps like Telegram.
The Future Of AI? 🤯 Hermes Agent Intro & Benefits #ai #agenticai #hermes
Hermes is an AI agent that improves over time by learning from user workflows, unlike most static AI tools. It automatically identifies recurring patterns, creates reusable skills, and applies them to similar future tasks without explicit training.
One Agent, Many Devices? 🧠 Full Hermes Agentic AI + Tailscale Workflow
This tutorial demonstrates how to access a single Hermes AI agent from multiple devices using Tailscale VPN and a remote dashboard setup, enabling unified session history, centralized file access, and continuous memory across laptop and desktop without fragmentation.
Safer Agentic AI? 🧠 Full Docker + Hermes Agent Tutorial (Desktop)
This tutorial demonstrates how to safely run AI agents like Hermes inside Docker containers to sandbox code execution and prevent unauthorized access to sensitive files on your computer. By configuring Docker volumes and mounting specific folders, users can give agents access to only the files they choose while keeping code execution isolated from the main system.