[How-To] Claude Dojo, Cú Chulainn
The Multi-Agent Orchestration Framework, The Visual Dojo, and The End of Terminal Hoarding
This week’s Office Hours was a dedicated Demo Day. We shifted focus from external market analysis to internal tool proliferation. As I mentioned on the call, “Internal tools are the best kind of tools” because they are built to solve the specific, bleeding edge friction points we face daily.
We showcased two major internal innovations: Cú Chulainn, a massive multi-agent orchestration framework built by our own Operations team, and Claude Dojo, a visual interface for managing parallel coding agents built by Rahul Subramaniam and the brightest minds in Trilogy Innovations.
Here is the breakdown of the architecture and workflows revealed during the session.
Claude Dojo: Visualizing Parallel Intelligence
Leonardo Gonzalez unveiled Claude Dojo, a brainchild of Rahul Subramaniam, designed to solve a specific developer pain point: Terminal Tab Hoarding.
The Context Compaction Problem
When working on complex features, developers often rely on a single agent instance. Over time, the context window fills up, and the agent begins to “compact” information, losing fidelity.
The Solution: Decompose the task. Run one agent for the frontend, one for the backend, and one for documentation.
The Friction: Managing 20+ terminal tabs is a nightmare. You lose track of which agent is doing what.
The Visual Interface
Claude Dojo uses a 3D Hexagonal Grid (built with Three.js) to visualize multiple running claude-code sessions simultaneously.
Status at a Glance: “Magic sparkles” indicate an agent is thinking; a static icon means it’s waiting for input.
Queueing: You can queue up commands (e.g. “When finished, write documentation”) so the agent keeps working while you step away.
Dangerously Skip Permissions: It supports flags to bypass the constant “Approve this edit” friction, allowing true autonomous loops.
“We’ve got terminal tab hoarding... now we’re getting into 20, 40 terminal tabs... Being able to see them all in one place visually is definitely a step up.”
Gamification vs. Utility
There is a growing trend to make agent management look like a video game (RTS style). While Claude Dojo looks slick, the long term value is high density information display, seeing the status of 10 parallel workstreams instantly.
Get Involved:
Join the Claude Dojo Google Chat Space to provide feedback and feature requests.
Cú Chulainn: The Multi-Agent Orchestration Framework
Presented by Fernando Perez and Ahmet Metin Birgili, Cú Chulainn (named after the Irish demigod hero) is not just a tool; it is a framework designed to decouple the task from the context.
The Architecture
Cú Chulainn is built on the agent tool paradigm (originally based on AutoGen) but adds a critical orchestration layer.
The Package: This container holds the context (product knowledge, rules).
The Adapter: This is the entry point (Jira, Slack, Kayako, BrainTrust) that introduces the specific task.
The Agents: Specialized workers (e.g. a “Reviewer”, a “Runner”) that execute the task using Python skills.
The “Andon Cord” for AI
One of the most robust features of Cú Chulainn is its error handling, dubbed the Andon Cord (borrowed from Toyota lean manufacturing).
The Problem: Single agents often hallucinate or enter “doom loops” where they repeat mistakes endlessly.
The Fix: The orchestration layer detects these loops. It stops the execution, alerts a human developer, and generates a log explaining exactly why the prompt failed, allowing for iterative improvement rather than silent failure.
“It handles errors gracefully... we have seen very good performance with multi-agentic when we have issues like... an agent goes into a doom loop... the cord goes like ‘hey wait wait wait... let’s call for human review.’”
Fernando Perez
Resources:
Deep Dive: I interviewed Fernando on the philosophy behind this in a mini-podcast here.
Documentation: Check out the Cú Chulainn Introductory Guide.
Try It: We have spun up a brand new sandbox environment for you to explore.
Problem Solution Log: The Internal Tooling Edition
We identified specific operational friction points and the internal software solutions deployed to fix them.
1. The “Doom Loop” Problem
Issue: Agents get stuck repeating errors or hallucinations in a single chat.
Solution: Cú Chulainn Andon Cord. An orchestration layer that monitors for loops, kills the process, and alerts a human developer to fix the prompt.
2. The Context Compaction Problem
Issue: Long sessions degrade in quality as the context window fills.
Solution: Parallel Agent Decomposition. Breaking a large task into smaller, parallel agents (Frontend/Backend/Docs) to keep context windows fresh.
3. The Terminal Fatigue Problem
Issue: Losing track of 20+ open terminal tabs for different agents.
Solution: Claude Dojo (cc-dojo). A web based visualizer that aggregates multiple CLI sessions into a single 3D dashboard.
4. The Integration Silo Problem
Issue: Agents living inside specific tools (only inside Jira, or only inside Slack).
Solution: Cú Chulainn Adapters. A modular entry point system that allows the same agent “package” to be triggered by email, Slack, Jira, or BrainTrust.
Next Steps
Try Claude Dojo: If you have
npminstalled, runnpx cc-dojoto test the visualizer locally.Explore Cú Chulainn: Use the sandbox link or ask Fernando for Admin access to break things.
Three Questions: We still have slots open for the “Three Questions” interview series. Reach out to me if you have a spiky point of view to share.
See the full meeting recording here