agentfdr

Detailed Description

agentfdr is a specialized debugging tool designed specifically for autonomous coding agents such as Claude Code. Its core purpose is to act as a local-first flight data recorder that captures and analyzes the detailed session transcripts of these agents, enabling developers to understand complex behaviors like looping, goal drift, or excessive token consumption. Unlike traditional debugging tools that require instrumentation, cloud services, or configuration, agentfdr operates entirely locally and reads the transcripts agents already produce, making it a zero-configuration, zero-cloud solution. This approach ensures privacy, ease of use, and minimal setup overhead, which is crucial for developers working with autonomous agents that generate extensive logs and token usage data. At the heart of agentfdr is its comprehensive timeline viewer, which presents the entire session on a single screen. This viewer details every turn the agent takes, including all tool calls made, how the context window is composed, and the output tokens generated. It also highlights prompt and compaction markers, allowing developers to visually track how the agent’s input and output evolve over time. Complementing this is an automatic anomaly detection system that flags common failure modes such as tool loops, error streaks, context bloat, token spikes, cache thrashing, file churn, and refusals. These alerts help developers quickly pinpoint problematic behaviors without manually sifting through logs. Further granularity is provided by the turn dissection panel, which breaks down each turn’s usage metrics, including detailed tool call statistics like duration and result size. This level of insight is invaluable for optimizing agent performance and diagnosing inefficiencies. Additionally, agentfdr offers a live watch mode that allows real-time monitoring of ongoing sessions, enabling developers to observe context growth and token usage as the agent operates. For continuous integration pipelines, the CI gate feature enforces limits on loops and token consumption, preventing runaway processes and reducing operational costs. Sharing insights is streamlined through the markdown autopsy feature, which renders the analysis into markdown format suitable for posting in issue trackers or Slack threads. This facilitates collaboration among development teams by making debugging data easy to communicate. Moreover, agentfdr aggregates usage metrics with per-model breakdowns and estimates costs for subscription metering, helping teams manage budgets and understand resource consumption. agentfdr is best suited for developers, AI researchers, and engineers who work with autonomous coding agents and need to debug complex, multi-turn interactions. It is particularly useful in scenarios where agents exhibit unexpected looping behavior, drift away from intended goals, or consume excessive tokens leading to high costs. By providing a detailed, turn-by-turn forensic analysis, agentfdr empowers users to engineer more reliable and efficient agent loops. Regarding pricing, agentfdr is an open-source tool available via npm and GitHub, which means it can be used freely without subscription fees. Its local-first design eliminates cloud costs and reduces dependency on external services. This makes it an attractive option for teams looking for a cost-effective debugging solution. Compared to alternatives, agentfdr stands out by focusing exclusively on autonomous coding agents and offering a zero-instrumentation, zero-cloud experience. Many other debugging tools require complex setup, cloud integration, or do not provide the same level of detailed session analysis and anomaly detection tailored to agent workflows. However, agentfdr’s local-first approach may limit collaborative real-time debugging features that cloud-based platforms offer. Notable limitations include its current specialization for agents like Claude Code, which means it may require adaptation to support other agent frameworks fully. Also, while it excels at post-session analysis, its live watch mode is primarily observational and may not support interactive debugging. Users should also be aware that interpreting the detailed metrics and anomaly reports requires some familiarity with autonomous agent behavior and token management. Overall, agentfdr is a powerful, niche tool that fills a critical gap in autonomous agent development by providing deep visibility into agent operations without the overhead of cloud services or instrumentation.