AGENTDEBUGX
5Demo Traces
5Failing Traces
117Findings Surfaced
19Failure Modes
About
Debugging agents requires structured evidence.
AgentDebugX turns raw trajectories into a readable failure object: what happened, why it matters, where the failure began, and how to verify the fix.
From failed run to root cause.
Follow planner decisions, tool calls, observations, state transitions, and final answers in one timeline with diagnosis evidence attached.
01Trace every model/tool/environment event with stable identifiers.
02Attach failure family, severity, confidence, and suspicious step evidence.
03Export the diagnosis as a case that future runs can replay.
Replay and preserve the fix.
Restart from checkpoints, compare alternate branches, export the evidence bundle, and save representative failures as regression cases.
5traces loaded
5failed runs
117findings
100.0%error rate
Framework
A closed loop for agent failures.
The workspace connects four activities that are usually scattered across logs, notebooks, spreadsheets, and ad-hoc reruns.
Agent
Debugging
Loop
Debugging
Loop
Load local JSONL, SQLite, or benchmark traces into a project workspace.
Group findings by failure family and highlight suspicious transitions.
Rerun from a checkpoint and compare original versus alternate trajectories.
Persist representative failures for evaluation suites and reports.
Applications
Where AgentDebugX fits.
Use it wherever agent behavior needs to be explained, reproduced, and turned into a durable evaluation artifact.
Benchmark teams
Failure-mode analysis
Analyze GAIA, AgentErrorBench, WebShop, and custom traces by root cause instead of pass rate alone.
Best for batch eval sweeps and failure-family review.
Product teams
Incident review
Move from a bad final answer to the upstream model, tool, or environment event that caused the run to drift.
Best for explaining production misses with concrete evidence.
Research teams
Reusable artifacts
Package failures into readable cases for papers, demos, ablations, and internal debugging reviews.
Best for demos, ablations, and shareable debugging artifacts.
Resources
Open the product surfaces.
The online homepage should route users into the workspace, overview, developer resources, and docs without hiding what the product does.
Teams
University of Illinois Urbana-Champaign
University of Illinois Urbana-Champaign
University of Toronto
University of Toronto
Google
Google
Stanford University
Stanford University
FAQ
What users need to know.
Common questions before you open the workspace.
No. The overview is a doorway; the core value is root-cause diagnosis and replay inside individual trajectories.
Sign in with Google and drop a trajectory .json, a JSON array, or a .jsonl file into your workspace. Known agent-log formats are converted automatically, and unrecognized JSON is mapped by an LLM converter on a best-effort basis.
Open any trace and run the Diagnose Pipeline: heuristic rules, an LLM judge, attribution, and DeepDebug root-cause localization — powered by the server's built-in LLM, no API key required.
A saved case includes the original trace, diagnosis evidence, replay comparison, and metadata to reproduce the issue.
Start with the failed run.
Open the workspace, choose a trace, and inspect the evidence behind the failure. The hosted version can later replace local storage with projects, teams, and uploads.
Launch App