I Can SEE EVERYTHING: Claude Code Hooks for Multi Agent Observability

Detailed Summary

Introduction (0:00 - 0:48)

The video highlights the critical challenge of observability when scaling beyond a single Claude Code agent, emphasizing that multi-agent systems quickly become unmanageable without proper tracking. It introduces a solution using Claude Code Hooks to send events to a client-server architecture, enabling comprehensive visibility into the actions of multiple agents simultaneously.

Multi-Agent Observability System Overview (0:48 - 2:57)

This section details the multi-agent observability system, which ingests events from all running Claude Code instances, regardless of their tasks (subtasks, planning, building, web scraping, deep research, calling MCP servers). The architecture consists of multiple agents firing Claude Code Hooks, which hit a Bun server. This server stores data in an SQLite database for persistence and simultaneously streams events via WebSockets to a client application, ensuring real-time tracking across various devices and agents. The system provides notifications when agents require input or complete tasks, demonstrating its interactive capabilities.

Application UI/UX and Features (2:57 - 11:48)

The presenter showcases the application's user interface, which includes filters for different codebases (e.g., "infinite agentic loop" and "cloud code hook multi-agent observability"), session IDs, and event types. A live activity pulse visually represents the real-time event flow from all agents across applications. Key features include color-coding for applications and session IDs for easy identification, and automatic summarization of events (especially pre-tool and post-tool use) using a small, fast model like Haiku. This summarization provides quick, at-a-glance insights into agent activities, complementing the ability to view full event payloads and chat transcripts for detailed debugging, particularly with stop events.

Codebase Architecture Breakdown (11:48 - 17:40)

The video delves into the technical implementation, starting with the Claude Code agent's hooks directory. It explains how a send event method is integrated into various hooks (pre-tool use, post-tool use, notification, stop, sub-agent stop, pre-compact). This method dynamically sends event data, including an optional summarize flag, to the server. The server, a simple Bun server, receives these events, inserts them into an SQLite database, and broadcasts them to WebSocket clients. The client-side, built with Vue.js, uses a use websocket composable to connect to the WebSocket, receive events, and update the UI. The architecture emphasizes a one-way data flow from agents to the server and then to the client, ensuring simplicity and efficiency.

The Importance of Observability (17:40 - 20:05)

The presenter stresses that without observability, generative AI technology and multi-agent systems are ineffective. The ability to measure, monitor, steer, correct, and control agentic systems is crucial for scaling work and understanding agent behavior. The use of small, fast models for summarization is highlighted as a cost-effective solution (less than 20 cents for thousands of events in testing). The video concludes by advocating for a shift from back-and-forth prompting to fully trusted, programmable agentic coding, positioning multi-agent observability as a key trend in the future of engineering.

Notable Quotes

"When it comes to multi-agent systems, observability is everything." — IndyDevDan

"With cloud code hooks sending events to a minimal client server architecture, you can see everything." — IndyDevDan

"This is the power of observability. As you scale up your agents, as you specialize and build agents into your workflows... you want to understand what they're doing." — IndyDevDan

"If you don't measure it, you can't improve it. If you don't monitor it, how will you know what's actually happening?" — IndyDevDan

"None of this generative AI technology matters if you have no idea what's going on." — IndyDevDan

"Prompting back and forth, one prompt at a time is not the way to engineer. It's a great place to start. It's a terrible place to finish." — IndyDevDan