DoVer: Automated Debugging for LLM Multi-Agent Systems via Intervention and Verification

Key points

• Paradigm shift: Traditional debugging of LLM multi-agent systems relies on log attribution — asking an LLM to identify the faulty agent and step from execution logs. This suffers from two flaws: hypotheses are never actually verified, and single-point attribution is ill-posed since multiple distinct interventions can independently fix the same failed task.
• Core principle (Do-then-Verify): DoVer generates fault hypotheses from execution logs, then verifies each by intervening at the suspected fault point (e.g., modifying a message or adjusting the plan), replaying execution from that checkpoint, and checking whether the failure is resolved. Debugging thus becomes grounded in repeatable experiments rather than uncertain labels.

Architecture

DoVer consists of six components:

1. Log Analyzer – preprocesses long execution logs and segments them into independent "trials" (typically delimited by re-planning events). 2. Hypothesis Generator – uses an LLM to produce candidate fault hypotheses (agent + step index + rationale). 3. Intervention Designer – converts each hypothesis into a minimal, executable intervention. DoVer focuses on orchestrator-level interventions: modifying messages sent to sub-agents, or updating the task plan (reordering, splitting, or replacing steps). 4. Intervention Executor – applies the intervention via a checkpoint-and-replay mechanism, restoring prior state and re-running the system from the intervention point (counterfactual execution). On AutoGen2, DoVer added checkpoint/replay interfaces to enable this. 5. Result Evaluator – judges outcomes by (a) task success, and (b) milestone progress: an LLM judge compares pre/post intervention execution against a list of key sub-goals extracted from the ideal solution path, computing a progress rate. 6. Controller – iterates the hypothesis–intervention–verification loop; hypotheses are labeled verified (success in at least one of two independent runs), partially verified (no success but significant progress gain), refuted, or inconclusive.

Experimental results

• Failure recovery: On Microsoft's Magnetic-One framework, DoVer converted 18% of failed tasks to successful on AssistantBench and 28% on GAIA; on AutoGen2/MathChat with GSMPlus it achieved 49%.
• Progress: On GAIA-Level-1, interventions yielded an average 16% milestone progress improvement even when tasks did not fully succeed.
• Hypothesis resolution: DoVer definitively verified or refuted 30%–60% of generated hypotheses.
• Baselines: Self-Refine and CRITIC recovered zero failure cases in DoVer's experiments; even prompt-optimized GPT-4o achieved only 24% accuracy in precise step-level log attribution.
• Inconclusive cases often arise when sub-agents lack the capability required to faithfully execute an intervention (e.g., a missing web-browsing tool). These cases are logged as improvement leads rather than discarded.

Qualitative insights

• In a GAIA task example, DoVer split the log into four trials, identified an invalid orchestrator instruction at step 53 (partially verified), then found a second independent fault in the trial-4 plan. Fixing both converted the task from failure to success, demonstrating multi-fault-point discovery.
• Partial verification (≥20% milestone gain without full success) flags hypotheses "close to the truth," giving developers concrete leads.

Limitations and outlook

Each intervention effectively re-runs the task, incurring compute cost, though this is offset by automation, parallel hypothesis testing, early termination on confirmed fixes, and integration with CI pipelines. DoVer points toward scalable, verifiable, experiment-driven debugging for multi-agent systems, with future work involving framework adaptation and combining automated interventions with human analysis for inconclusive cases.

Case example

DoVer's intervention at step 53 of trial 2 (correcting the orchestrator's instruction to click the right button) achieved key milestones but not full success; a second intervention on trial 4's plan (replacing direct scrolling with calendar-based navigation) completed the task — illustrating how DoVer locates and repairs multiple independent fault points within one task.