The UC Berkeley Sky Computing Lab has officially released Agency, an open-source platform designed to bring visibility to the black box of AI agent execution. As agents increasingly handle complex, parallel tasks involving multiple tool invocations, traditional model traces have proven insufficient for understanding latency bottlenecks or system-level waits. Agency aims to solve this by connecting LLM requests, tool calls, and underlying system activity into a single, navigable interface.

Visibility for the Obscure

The core philosophy of Agency is described by the developers as 'htop + gdb for AI agents.' This analogy highlights the tool's dual focus on real-time monitoring and deep-dive debugging. Users can watch agents execute in real time, inspecting model requests, outputs, failures, and wait states across multiple concurrent agents. The dashboard allows for direct control, enabling users to pause, resume, or stop supported executions on the fly. Once an execution finishes, the live view transitions into a historical record, allowing for post-mortem analysis of exactly what happened during the run.

Beyond Simple Traces

Agency distinguishes itself with a Critical Path Profiler that maps where wall-clock time is actually spent. This feature traces dependencies across agents to identify what work determines completion time, clearly distinguishing between parallel execution and blocking waits. For deeper analysis, the Detailed Profiler exposes metrics such as LLM latency, token usage, tool durations, errors, retries, and system resources including CPU, memory, I/O, and optional GPU metrics. The platform emphasizes factual accuracy, ensuring that missing measurements remain missing rather than being filled in with guesses.

Model and Harness Agnostic

The platform is explicitly model and harness agnostic, supporting a wide range of popular coding agents including Claude Code, Codex, OpenCode, and Agency's own native harness. It integrates with major model backends such as OpenAI, Anthropic, Amazon Bedrock, and local models via vLLM. The architecture provides a sandboxed execution environment using Docker or Podman on Linux x86-64 systems, requiring Python 3.12+ and Git. While integration with existing CLI sessions for tools like Claude Code is currently experimental, Python-orchestrated workflows offer the most complete experience for now.

Key Takeaways

  • Agency is an open-source platform from UC Berkeley Sky Computing Lab for debugging and profiling AI agents.
  • It offers real-time execution visualization, critical path profiling, and detailed system metrics (CPU, GPU, I/O).
  • The tool supports multiple agent harnesses (Claude Code, Codex, OpenCode) and model backends (OpenAI, Anthropic, vLLM).
  • Installation requires Linux x86-64, Python 3.12+, Docker/Podman, and a valid API key for the chosen model provider.
  • The dashboard runs locally on localhost:7860 and binds to 0.0.0.0 by default, requiring access controls for network exposure.

The Bottom Line

If you are running multi-agent systems and still relying on raw stdout logs to find bottlenecks, you are flying blind. Agency provides the necessary instrumentation to turn chaotic agent swarms into observable, debuggable systems.