The rise of AI‑powered browser automation, RPA scripts, and crawler frameworks has unlocked unprecedented productivity, but it has also introduced a new class of elusive failures that are difficult to trace and remediate.
Agent Failure Doctor arrives as a purpose‑built, local‑first diagnostic toolkit that tackles these challenges head‑on.
By ingesting raw artifacts such as trace.zip, error logs, console output, network JSON, probe reports, screenshots, OCR evidence, and user‑provided descriptions, the tool reconstructs a detailed picture of what went wrong during a run.
This approach sidesteps the need for external telemetry or cloud‑based correlation, keeping sensitive data on the developer’s machine while still delivering actionable insights.
Traditional debugging of Playwright, Selenium, or custom RPA bots often involves sifting through fragmented logs, reproducing flaky UI states, and guessing root causes based on limited evidence.
Agent Failure Doctor flips this model by enforcing an evidence‑bound reasoning process: it prioritizes concrete artifacts over speculation and will explicitly label a case as “insufficient_evidence” when the data does not support a confident conclusion.
This disciplined stance reduces the risk of chasing false leads and ensures that remediation efforts are grounded in observable facts.
The tool accepts a remarkably broad set of inputs, reflecting the heterogeneous nature of modern automation stacks.
A typical invocation might point to a trace.zip generated by Playwright, an error.log from a crawler, console.txt capturing stdout/stderr, network.json detailing HTTP exchanges, probe_report.json from synthetic monitoring, screenshot metadata, a user_description.txt that explains the intended scenario, visual_run artifacts, and any OCR or document evidence collected during execution.
By normalizing these diverse formats into a common internal structure, the diagnostic engine can cross‑reference timing, DOM state, network behavior, and visual changes to pinpoint inconsistencies that would be invisible when looking at each artifact in isolation.
Once the evidence is gathered, Agent Failure Doctor produces a multi‑faceted output designed to move a team from confusion to concrete action.
The core report includes a concise conclusion summarizing the likely failure mode, an evidence section that cites the specific artifacts supporting that conclusion, a “why” explanation that walks through the causal chain, a recommended next action, and repair suggestions that can be turned into code changes.