try1004/flowlens-agentops

FlowLens AgentOps is a diagnostics and monitoring layer for systems that run multiple AI agents together. It is built on top of a project called DeerFlow, which supplies the underlying agent runtime, tools, memory, and sandbox, while FlowLens adds the ability to watch those runs, understand what happened during them, and figure out why something failed. The core problem it addresses is that when a multi-agent run fails, a saved checkpoint can restore the system to a working state, but it does not tell you why the failure happened, and a final error message loses all the decisions and steps that led up to it. FlowLens records a lightweight, ongoing log of runtime events and turns that log into a normalized timeline of everything that happened across models, tools, subagents, and other parts of the system. It also produces run metrics such as duration, token usage, and error counts, and it applies rule based logic to point at a likely cause of failure, complete with supporting evidence, other contributing factors, a confidence level, and a suggested next step. There is also an optional AI generated explanation layer, though the rule based result itself cannot be changed by it. The product includes a shared dashboard with several views: a run explorer for searching past runs, an execution timeline you can filter by phase, health metrics, and a root cause view. It also ships seven ready made replay scenarios, such as a tool error or a subagent timeout, that let you explore the dashboard with no API key, no account, and no backend required. A separate connected mode reads real data from a live DeerFlow system instead. Getting started with the no key replay mode involves cloning the repository and running it either through a Node package manager or through Docker, after which a demo dashboard opens in your browser. Running the fully connected version requires a working DeerFlow setup, including authentication and at least one configured AI model provider. The diagnostics path is explicitly read only: it does not change checkpoints, prompts, or the underlying run records it inspects. The README states its test suite passes across backend and frontend checks and includes a secret scan. FlowLens is released under the MIT License, the same license used by the DeerFlow project it builds on.