AgentDebugX

A local-first debugging framework for agentic AI systems: diagnose failures, attribute root causes, recover with evidence, and validate fixes through reruns.

AgentDebugX turns failed agent runs into structured, auditable debugging artifacts. It ingests a live or exported trajectory, detects visible failure signals, attributes them to responsible steps or agents, proposes recovery actions, and prepares controlled reruns so fixes can be validated instead of guessed.

The project is designed for researchers and engineers building complex LLM agents: multi-agent systems, tool-using agents, computer-use agents, benchmark runners, and local agent development workflows. AgentDebugX is local-first by default: traces stay on your machine, sharing is opt-in, and recovery proposals carry explicit policy and approval metadata into the Rerun boundary.

📰 News

• 🔌 2026-08-25 — Released dsh-agentdebugx v0.1.0, the AgentDebugX plugin for DeepSeek Harness.
• 📄 2026-07-31 — Released CUADebug, our framework for diagnosing and repairing computer-use agent failures.
• 📄 2026-07-21 — Released the AgentDebugX paper, presenting our open-source toolkit for failure observability, attribution, recovery, and rerun in LLM agents.
• 📦 2026-05-16 — Released AgentDebugX on PyPI.
• 📄 2025-09-29 — Released Where LLM Agents Fail and How They Can Learn From Failures, introducing AgentErrorTaxonomy, AgentErrorBench, and AgentDebug.

System Overview

AgentDebugX follows the two-stage loop used by the project paper:

Diagnose = Detect -> Attribute -> Recover
Rerun    = checkpoint -> retry directive -> branch execution -> evaluation

Diagnose explains what failed and why. Rerun tests whether the proposed recovery actually improves the agent behavior.

Why AgentDebugX

Tracing tools show what happened. AgentDebugX focuses on the debugging step that usually comes next:

• Which earlier decision caused the visible failure?
• Which agent, tool call, memory read, handoff, or GUI action was responsible?
• What evidence supports that diagnosis?
• What concrete recovery should be tried?
• Did the rerun branch improve the outcome?

The output is a portable diagnostic report that can be inspected in a local UI, used by a CLI workflow, stored in an Error Hub bundle, or invoked from an agentic skill.

Core Capabilities

• Portable trace schema: framework-agnostic trajectory, event, finding, and diagnostic report models.
• Ingest adapters: normalize raw JSON, LangGraph, CrewAI, OpenAI Agents SDK, OpenTelemetry, GAIA/Open Deep Research, OSWorld, and other exported traces.
• Detect: deterministic analyzers, manifest-backed rule packs, LLM judge mode, GUI-aware signals, and taxonomy induction support.
• Attribute: heuristic attribution, all-at-once analysis, step-by-step localization, binary search, counterfactual attribution, MOE localization, and DeepDebug.
• Recover: Reflexion, CRITIC, Self-Refine, AutoManual, DeepDebug recovery, and saga rollback style strategies.
• Rerun: three explicit modes for plan/export only, labeled simulation, or observed execution in an application-owned process or persistent HTTP runner.
• Local inspection UI: no-build FastAPI dashboard for traces, reports, before/after CUA visuals, debugger discussions, saved cases, debug branches, and rerun-from-event workflows.
• Error Hub: scrubbed, shareable failure bundles for regression tests, benchmark corpora, and team debugging memory.
• Agent integrations: generate host-runtime assets such as debugging skills for external agent tools.

Install

pip install agentdebugx

Optional extras:

pip install "agentdebugx[ui]"             # local FastAPI dashboard
pip install "agentdebugx[langgraph]"      # LangGraph adapter
pip install "agentdebugx[crewai]"         # CrewAI adapter
pip install "agentdebugx[openai-agents]"  # OpenAI Agents SDK adapter
pip install "agentdebugx[otel]"           # OpenTelemetry ingest
pip install "agentdebugx[gui]"            # screenshot decoding for GUI RCA
pip install "agentdebugx[all]"            # all optional integrations

Computer-use / OSWorld GUI root-cause analysis (agentdebug.gui) ships with the core install and needs no extra. The gui extra only adds pillow, which the RCA tools use to decode screenshots. Two heavier layers of the same package sit behind their own extras: gui-memory for the lesson/episodic memory stack, and gui-app for the provider adapters, the batch pipeline (python -m agentdebug.gui) and the Streamlit annotation app.

The package is installed as agentdebugx and imported as agentdebug:

import agentdebug

DeepSeek Harness Plugin

AgentDebugX is also available as the dsh-agentdebugx plugin for DeepSeek Harness. It diagnoses current and saved Harness trajectories and starts the Python bridge and local dashboard only when they are needed.

pip install "agentdebugx[ui]>=0.3.1,<0.4"
dsh plugin --profile web add dsh-agentdebugx

See the plugin documentation for configuration, commands, saved-session discovery, and deep diagnosis.

Quick Start: Python API

Record a trajectory and analyze it locally:

from agentdebug import AgentDebug, EventType

debugger = AgentDebug()

with debugger.trace(
    goal="Book a refundable NYC to SFO flight",
    framework="my-agent",
) as trace:
    trace.record(
        EventType.PLAN,
        agent_name="planner",
        step_index=1,
        output="Search for the cheapest fares.",
    )
    trace.record(
        EventType.TOOL_RESULT,
        agent_name="browser",
        step_index=3,
        error="Checkout failed: refund_policy is required.",
    )

    report = trace.analyze()

print(report.summary)
for finding in report.findings:
    print(finding.failure_mode.mode_id, finding.step_index, finding.evidence)

The report localizes the responsible upstream step rather than only reporting the final visible error.

Quick Start: CLI

The CLI supports the complete Diagnose -> Rerun workflow. The web console is optional and is not required for trace conversion, diagnosis, attribution, recovery planning, or rerun preparation.

1. Normalize an external trace

AgentDebugX can auto-detect common JSON and JSONL exports:

agentdebug ingest raw_trace.json --format auto --out trace.json

Use --format when the source is known, for example messages, openai_agents_spans, crewai_events, langgraph_callbacks, claude_code, or osworld.

Process a directory of independent JSON files or every non-empty line in a JSONL dataset:

agentdebug batch ingest AgentProcessBench/gaia_dev/test.jsonl \
  --out-dir data/agentprocessbench/gaia_dev

Batch diagnosis normalizes each record and writes independently rerunnable trajectories and reports:

agentdebug batch diagnose AgentProcessBench/gaia_dev/test.jsonl \
  --mode judge \
  --attributor all-at-once \
  --recovery self-refine \
  --out-dir runs/agentprocessbench/gaia_dev

Every batch writes batch-summary.json. Invalid records are isolated and do not discard successful outputs; a partially failed CLI batch exits with code 3.

2. Run a fully local diagnosis

The deterministic pipeline does not require an API key:

agentdebug diagnose trace.json \
  --mode heuristic \
  --attributor heuristic \
  --recovery reflexion \
  --out report.json

Render the same diagnosis as a cascade-oriented traceback:

agentdebug diagnose trace.json \
  --mode heuristic \
  --attributor heuristic \
  --recovery reflexion \
  --traceback

3. Enable LLM-backed diagnosis

Save an OpenAI-compatible endpoint once:

agentdebug config set-llm \
  --base-url "https://<openai-compatible-host>/v1" \
  --api-key "<secret>" \
  --model "<model>"

Then select the diagnosis, attribution, and recovery implementations explicitly:

agentdebug diagnose trace.json \
  --mode judge \
  --attributor all-at-once \
  --recovery self-refine \
  --out report.json

For difficult multi-step or ambiguous failures, DeepDebug runs the complete diagnosis workflow and automatically packages its evidence-backed fix as a standard retry directive:

agentdebug diagnose trace.json \
  --mode deepdebug \
  --out report.json

--recovery deepdebug can select this packaging explicitly. Existing scripts that use --attributor none --recovery none remain compatible; explicit --recovery none disables the standard recovery payload.

Environment variables can be used instead of saved configuration:

export AGENTDEBUG_LLM_BASE_URL="https://<openai-compatible-host>/v1"
export AGENTDEBUG_LLM_API_KEY="<secret>"
export AGENTDEBUG_LLM_MODEL="<model>"

Use agentdebug config show to inspect masked configuration and agentdebug config doctor to test the configured endpoint.

4. Execute the Rerun stage

For repeated, Docker, or remote reruns, keep the application's complete Agent environment running as an HTTP runner service. Implement a project callback, then start and configure it once:

agentdebug runner serve my_project.runner:run_agent \
  --name my-agent \
  --framework langgraph \
  --host 0.0.0.0 \
  --port 8765 \
  --token-env MY_RUNNER_TOKEN

agentdebug config set-runner my-agent \
  --url http://127.0.0.1:8765 \
  --token-env MY_RUNNER_TOKEN \
  --default

agentdebug config doctor-runner my-agent

Then run the original agent framework from the beginning of the task:

agentdebug rerun report.json \
  --trajectory trace.json \
  --out rerun.live.json

To branch from a specific trajectory event, pass its 1-based event number:

agentdebug rerun report.json \
  --trajectory trace.json \
  --start-event 4 \
  --out rerun.from-event.json

--start-event N resolves the Nth event to its stable event ID and sends a from_event checkpoint to plan, simulation, and live runner modes. The selected runner must support restoring or continuing from event checkpoints.

The service owns the framework, real model, tools, credentials, environment, job lifecycle, and trajectory recorder. A chat-completions URL alone is not an Agent environment. Submissions are idempotent, transient failures use bounded retries, and unfinished remote jobs are cancelled best-effort. See the runner specification.

For local scripts and CI, the process compatibility transport remains available:

agentdebug rerun report.json \
  --trajectory trace.json \
  --runner-command "python path/to/project_rerun_runner.py" \
  --out rerun.json

Use --plan-only for trajectory-only uploads; the plan reports why real execution is unavailable and which runtime capabilities are missing.

Export the same request as a pending actor task dataset when another system will perform the rollout:

agentdebug rerun report.json \
  --trajectory trace.json \
  --plan-only \
  --actor-task-format jsonl \
  --out rerun-tasks.jsonl

Parquet is also supported with --actor-task-format parquet after installing pyarrow. These rows contain actor inputs and provenance, not responses or training labels. See the actor task specification.

For prompt experiments only, --simulate enables the previous LLM-generated trajectory flow. It returns a workflow JSON with status=simulated, a validated hypothetical_trajectory, and model-generated events explicitly marked as simulated. It executes no tools and is not evidence that the recovery fixed the task. See the simulation specification.

5. Work with stored traces

The CLI can query SQLite or JSONL stores created by instrumented runs or the local console:

agentdebug list --store-sqlite .agentdebug/traces.sqlite
agentdebug show <trace-id> --store-sqlite .agentdebug/traces.sqlite
agentdebug diagnose <trace-id> \
  --store-sqlite .agentdebug/traces.sqlite \
  --mode heuristic \
  --attributor heuristic \
  --recovery reflexion

6. Package and integrate debugging workflows

Publish a scrubbed failure bundle to a local Error Hub:

agentdebug hub push <trace-id> \
  --store-sqlite .agentdebug/traces.sqlite \
  --to local:./agentdebug-hub

Generate a debugging skill for a supported host runtime:

agentdebug integrations skill --platform claude --target .claude/skills

Optional: launch the local console

Install the UI extra only when a visual inspection workflow is useful:

pip install "agentdebugx[ui]"
agentdebug serve \
  --store-sqlite .agentdebug/traces.sqlite \
  --host 127.0.0.1 \
  --port 7777

CLI Reference

Run agentdebug <command> --help for version-specific flags.

Architecture

The repository mirrors the paper-level workflow:

src/agentdebug/schema/       portable trajectory, event, report, and taxonomy contracts
src/agentdebug/runtime/      storage, LLM clients, event bus, and plugin registry
src/agentdebug/ingest/       live capture APIs and offline trace importers
src/agentdebug/diagnose/     Detect -> Attribute -> Recover pipeline
src/agentdebug/rerun/        rerun plans, requests, branch comparison, and executors
src/agentdebug/inspect/      traceback renderer and local inspection UI
src/agentdebug/hub/          scrubbed failure bundle packaging and backends
src/agentdebug/integrations/ host skill and runtime integration generators
src/agentdebug/gui/          computer-use / OSWorld GUI root-cause analysis
examples/                    runnable examples and demo traces
docs/                        architecture, schema, and project assets

Detailed references:

• Architecture
• Trace schema
• System overview PDF

Component Model

Diagnose components use manifest-backed discovery:

• Detect components and rule packs declare metadata under src/agentdebug/diagnose/component_manifests/detect/ and src/agentdebug/diagnose/detect/rules/packs/.
• Attribute components declare metadata under src/agentdebug/diagnose/component_manifests/attribute/.
• Recover components declare metadata under src/agentdebug/diagnose/component_manifests/recover/.

The shared registry exposes:

from agentdebug.diagnose import list_components, load_component

for component in list_components():
    print(component.id, component.stage, component.capabilities)

This keeps the implementation extensible without turning the CLI or UI into business-logic containers.

Local UI

The inspection UI is a local FastAPI application with a no-build HTML/CSS/JS frontend. It is intentionally a surface layer:

• routes live in inspect/ui/routes.py
• rendering lives in inspect/ui/views.py
• UI-facing services live in inspect/ui/services.py
• local case and branch stores live in inspect/ui/branch_store.py
• inspect/ui/server.py remains a compatibility import path

Launch from the CLI

Install the optional UI dependencies, then point the server at an existing AgentDebugX trace store:

pip install "agentdebugx[ui]"

agentdebug serve \
  --store-sqlite .agentdebug/traces.sqlite \
  --host 127.0.0.1 \
  --port 7777

Open http://127.0.0.1:7777 in a browser. For a JSONL store, replace --store-sqlite with --store-jsonl .agentdebug/traces.jsonl. Keep the default loopback host unless the UI is deployed behind appropriate authentication and transport security. Place native trajectory and diagnostic-report JSON files under .agentdebug/imports/, then use Sync imports in the workspace to import new or changed files. Set AGENTDEBUG_IMPORT_DIR to use another server-owned directory.

The UI can inspect traces, save typical error cases, prepare debug continuations, and invoke a server-controlled live runner. Rerun Composer is opened from a selected event and uses that event as its checkpoint. Set AGENTDEBUG_RUNNER_URL for the preferred persistent HTTP transport or AGENTDEBUG_RERUN_COMMAND for process compatibility. The selected runner must advertise or implement from_event checkpoint support. The browser does not accept or persist runner commands or bearer tokens; LLM API keys configured in the local UI are retained only for the current browser tab.

OSWorld trajectories with locally available screenshot artifacts open in the read-only Visual view by default. Use the Trace / Visual control to switch without changing the selected event; the choice is remembered per trace for the current browser tab. Visual compares the selected action's Before state (an explicit input image or the immediately preceding event result) with all After images attached to the selected event; it never changes timeline selection or guesses across missing steps. Screenshots are served only through trace/event artifact IDs, and only when the resolved image remains inside the trajectory's recorded metadata.source_dir.

Discuss with Debugger is available for every normalized trace format, not only CUA. Discussions are persisted locally, pinned to a report snapshot, cite canonical event IDs, and may produce an exportable report-revision draft. Discussion tools are read-only and drafts never overwrite stored diagnostic reports. The separate Streamlit app remains the tool for annotation, multi-reviewer assignment, and accuracy workflows.

Privacy and Safety

AgentDebugX is local-first:

• Trace capture and diagnosis run locally unless you explicitly configure an external LLM endpoint.
• Error Hub publishing is opt-in.
• Bundle scrubbing is available before sharing traces.
• Recovery is suggest-only. External execution belongs to Rerun and requires an explicitly configured executor. Recovery approval fields are auditable metadata; deployments must enforce their own authorization policy before dispatch.

Diagnostic findings are hypotheses with evidence and provenance, not ground truth. LLM Judge reports retain the model's self-reported confidence; Heuristic and DeepDebug reports omit uncalibrated confidence values. Configure retention, access control, and redaction before collecting production traces.

Examples

The examples/ directory contains runnable scripts and demo artifacts:

• basic_usage.py
• multi_agent_cascade.py
• langgraph/
• crewai/
• autogen_roundrobin_deepdebug.py
• taxonomy_induction_demo.py
• http_agent_runner.py
• live_rerun_runner.py
• claude_skill_integration/
• debug_skills/

Development

Run the test suite:

python -m pytest tests -q

Run the enforced branch-coverage baseline:

python -m pytest tests -q --cov=agentdebug --cov-branch --cov-fail-under=40

Compile-check the package:

python -m compileall -q src/agentdebug tests

Build artifacts under dist/ should not be committed. Generate them only for release workflows.

See CONTRIBUTING.md for test organization, the optional GUI test suite, quality checks, and pull request expectations.

Citation

@article{agentdebug2025,
  title={Where LLM Agents Fail and How They Can Learn From Failures},
  author={Zhu, Kunlun and Liu, Zijia and Li, Bingxuan and Tian, Muxin and Yang Yingxuan and Zhang, Jiaxun and others},
  journal={arXiv preprint arXiv:2509.25370},
  year={2025}
}

@misc{zhu2026agentdebugxopensourcetoolkitfailure,
      title={AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents}, 
      author={Kunlun Zhu and Xuyan Ye and Zhiguang Han and Yuchen Zhao and Bingxuan Li and Weijia Zhang and Muxin Tian and Xiangru Tang and Pan Lu and James Zou and Jiaxuan You and Heng Ji},
      year={2026},
      eprint={2607.18754},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.18754}, 
}

@misc{zhang2026cuadebugdiagnosingrepairingcomputeruse,
      title={CUADebug: Diagnosing and Repairing Computer-Use Agent Failures},
      author={Weijia Zhang and Kunlun Zhu and Zeyi Liu and Yinting Chen and Tianyi Ma and Jiateng Liu and Jiaxun Zhang and Bingxuan Li and Xiangru Tang and Heng Ji and Jiaxuan You},
      year={2026},
      eprint={2608.02643},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2608.02643},
}

License

MIT. See LICENSE.