Agents Observe: Real-Time Claude Code Monitoring

Agents Observe is a real-time dashboard for monitoring Claude Code agents and their subagents. It uses hooks to capture detailed tool calls and execution timelines, providing transparency into autonomous AI workflows. The tool is available as a Claude plugin or a standalone Docker application, featuring a React UI for searching and filtering session data.

Key Points

• Provides real-time visualization of multi-agent hierarchies and subagent activities that are usually opaque in the terminal.
• Captures 'ground truth' data by logging every tool call, command execution, and file interaction via hooks rather than standard telemetry.
• Features a comprehensive React dashboard with live WebSocket updates, powerful filtering, and historical session search.
• Supports flexible deployment options, including a simple Claude plugin installation or a standalone Docker-based setup.
• Enables deep debugging by allowing users to expand specific events to see full payloads and trace the exact sequence of agent actions.

Sentiment

The community is broadly positive and enthusiastic about the tool, seeing it as addressing a genuine gap in multi-agent observability. However, several commenters raise the deeper concern that visibility into tool calls alone does not solve the problem of agents producing bad or misleadingly optimistic outputs. The tone is constructive and technically engaged rather than critical.

In Agreement

• Blocking hooks severely degrade agent performance; background hooks are essential for multi-agent setups
• The opacity of multi-agent workflows is a major pain point that observability dashboards directly address
• Docker-based service architecture is a smart pattern for security and process isolation
• Sub-agent teams with fresh context windows are significantly lighter on token usage
• The tool fills a critical gap between tailing logs manually and having real-time structured visibility

Opposed

• Event stream logging shows what tools were called but cannot determine whether an agent's self-reported outcome matches reality — the fundamental trust problem remains unsolved
• Many comments on Show HN posts for AI-coded projects appear to be bot-generated, undermining authentic discussion
• A simpler architecture without a server component might suffice if data is just going to SQLite