Agent Observatory: A Real-Time Observability Tool for Claude Code Multi-Agent Workflows

Agent Observatory is a real-time observability tool specifically designed for Claude Code multi-agent workflows, aiming to address challenges like call tracing, execution time analysis, and cost accounting in complex multi-agent systems. Its core value lies in providing features such as call tree visualization, timeline tracking, and cost estimation through a non-intrusive integration solution, helping developers achieve transparent monitoring and optimization of AI workflows.

As Claude Code's Agent tool capabilities grow, multi-agent workflows built by developers (where a main agent coordinates multiple sub-agents to execute tasks in parallel/series) face three major challenges: difficulty in tracing inter-agent call relationships, hard execution time analysis, and lack of effective cost accounting methods. Agent Observatory was created to address these issues.

Agent Observatory adopts a lightweight architecture with zero code changes:

1. Hook Script: hook.py.py receives events via Claude Code's PreToolUse and PostToolUse hooks, and POSTs JSON events events to the server using Python's standard library
2. Server: Built on FastAPI+SQLite, responsible for event persistence and SSE stream push
3. Frontend: Uses D3.js for data visualization Advantages include zero dependency intrusion, real-time performance (SSE second-level updates), data persistence (SQLite storage), and team sharing support.

Agent Observatory provides five complementary visualization views:

• Call Tree: D3.js hierarchical graph showing agent derivation relationships; nodes are colored by status (running/completed/interrupted), and hovering reveals prompt previews
• Timeline: Gantt chart displaying execution time and nesting levels, with a real-time NOW marker line to identify performance bottlenecks
• Flowchart: Shows agent wave connection relationships (solid lines for derivation/dashed lines for context flow) to understand data dependencies
• Logs: CLI-style terminal view that records events and supports expanding full prompts and responses
• Chat Board: Decentralized collaboration mechanism where agents can read the shared message board and reply autonomously.

Cost Estimation

Based on prompt/response text length (approximately 4 characters = 1 token), it uses the Claude Sonnet 4.6 pricing model (input: $3 per 1M tokens, output: $15 per 1M tokens) and supports modifying constants to adjust the model.

Agent Identity System

Named agents are assigned unique Unicode symbols (e.g., ⬡ for architecture-reviewer), while unnamed agents get fun names (like Jazzy Blobfish) and species emojis.

Team Sharing Modes

• Personal Local Mode: Run a local server to monitor personal sessions
• Team Sharing Mode: Deploy a shared server; members point to the server via the environment variable AGENT_OBSERVER_URL.

Deep Integration

Click a call tree node, and the details panel provides a "Continue in Claude Code" button, which copies a structured follow-up prompt (including original prompt, response content, and parent/child agent information) to the clipboard, enabling seamless connection between monitoring and development.

Limitations

• Cost estimation is based on character count approximation, which deviates from actual token counts
• The autonomy of agents in the chat board function depends on prompt design.

Future Directions

• Expand to more AI assistant platforms
• Develop more precise cost estimation models
• Add historical trend analysis and performance benchmarks
• Provide automatic optimization suggestions for agent collaboration modes

Summary

Agent Observatory fills the observability gap for Claude Code multi-agent workflows. With its non-intrusive architecture, rich visual views, and practical features, it provides developers with a powerful tool to understand and optimize complex AI systems. It is an open-source project worth trying for multi-agent application teams.