AgentDbg is a critical utility in the agent developer stack, focusing specifically on the observability and debugging layer. As agents move from simple chatbots to autonomous systems that execute code and call external APIs, the ability to trace their decision-making process becomes essential. AgentDbg provides this visibility locally, allowing for rapid iteration during the development phase.
The tool is particularly relevant for its focus on agent-specific failure modes, such as infinite tool-calling loops. By offering native support for frameworks like LangChain and the OpenAI Agents SDK, it integrates into the existing workflows of most agent builders. It champions a privacy-first approach to AI development, ensuring that proprietary prompts and sensitive tool data never leave the local environment during the debugging process.
Debugging an LLM-based agent is often an exercise in frustration. Unlike traditional software, where a stack trace usually points to a logical error, agent failures are often subtle. An agent might call a tool with the wrong arguments, get caught in a recursive loop that drains an API budget, or hallucinate a state change that never occurred. This is the friction point AgentDbg aims to solve.
AgentDbg is a local-first debugger designed specifically for the non-deterministic nature of agentic workflows. It provides developers with a structured timeline of every action an agent takes, including LLM prompts, tool invocations, and internal state updates. By capturing these events as they happen, the tool allows developers to step through a run as if they were using a traditional debugger, but with the added context required for generative AI.
The technical implementation relies on a simple @trace decorator. Once added to an agent's code, AgentDbg begins recording structured runs into local JSONL files. This approach is a deliberate departure from the growing category of cloud-based AI observability platforms. While those services offer centralized dashboards and team collaboration, they also require sending sensitive prompt data and tool outputs to a third-party server. AgentDbg keeps everything on the developer's machine, making it a viable choice for projects with strict data privacy requirements.
Beyond simple logging, the tool addresses the "token-wasting loop" problem. Because agents are autonomous, they can occasionally enter a state where they repeatedly call the same tool or re-try the same failing prompt. AgentDbg includes loop warnings and can even auto-stop agents that appear to be stuck. This is a practical utility for managing the costs associated with high-token-usage models like GPT-4o or Claude 3.5 Sonnet.
The project is open-source and supports popular frameworks including LangChain and the OpenAI Agents SDK. It is founded by Zafar Takhirov, who has positioned the tool as a developer utility rather than a massive enterprise platform. The interface is a clean, timeline-based UI that reveals the logic behind agent failures. Users can see the exact prompt that led to a specific tool call, the raw output from that tool, and how the agent integrated that information into its next step.
In the current market, AgentDbg competes indirectly with platforms like LangSmith or Arize Phoenix. However, its focus is narrower and more developer-centric. It does not try to be an evaluation harness or a production monitoring suite for thousands of concurrent users. Instead, it is the tool you reach for when you are at the terminal, your agent is doing something unexpected, and you need to see the raw state of the system. It is a piece of infrastructure for the build phase of the AI lifecycle, emphasizing local control and immediate feedback over long-term analytics.
A local-first, timeline-based debugger for inspecting AI agent LLM calls, tool calls, and state changes.
AgentDbg is hiring.