BLADE audits multi-agent handoffs to localize failure — Handoff auditing — What does it mean?
The news. On October 13, 2026, researchers at the University of Connecticut posted BLADE (Blame Localization via Audited Dependency Evidence) to arXiv. It needs no training and no reference answer, and it was tested on 1,228 failed runs from the Who&When and Who&When Pro benchmarks, which cover several multi-agent architectures and both natural and injected failures. Under one shared decision model, it had the best or tied-best step accuracy of eight methods on all five benchmark subsets. Read the paper →
Picture three nursing shifts. The lab note says the two patients came from Braintree and Honolulu. Nurse B reads it at the handover and writes Honolulu and Quincy into the shift notes, and the night shift copies those notes into the final chart. In the morning the chart is wrong. The chart is where the mistake shows, but Nurse B's handover is where it was made, and the night shift did nothing wrong except trust that handover. This is the paper's own example with agents in place of nurses: a code-execution step returned the correct pair of birthplaces of U.S. presidents, and the next agent misread it.
Now suppose all you have is the list of who briefed whom. That list says the final chart depends on Nurse B, and Nurse B depends on the lab note. A who-briefed-whom record cannot tell you which link broke, because a faithful copy of a wrong note and a misreading of a right note leave the same arrow. That is the gap between dependency tracing and auditing use, and it is why errors compound silently across a team: each agent's output looks reasonable given what it received.
BLADE works in three passes. First, it reconstructs dependencies as probabilities. For every event, Jev answers four multiple-choice questions (which earlier event does this one depend on, repeat, contradict, or repair), and any link with probability ≥ 0.3 is kept as a weighted edge. These are plausible links, not proven causes.
Second, it audits only the handoffs that matter. For each event it takes at most the 2 sources with the highest depends-on probability, and asks Jev how the receiving decision used each one, against the requirements extracted from the task text. The answer is one of nine labels in three groups: kept (faithful use, no real use, or not enough evidence), breach (an error introduced here), and inherited or repaired (an existing error carried forward, or fixed without a new one). The dependency edges and the non-kept findings together form a blame graph.
Third, a global diagnosis reads the whole trajectory plus the blame graph and picks the decisive step, and every step stays a candidate. The audit is evidence, not a verdict: a local breach can be fixed later or not matter, and the decisive error can sit at a step that was never audited. Leaving the final call open to every step is what the paper calls a defeasible diagnosis, meaning one that later evidence can overrule. In incident terms, it is root-cause discipline: collect evidence at each boundary, then decide.
| Method (same Jev model, no reference answer) | What it looks at | Step accuracy, macro avg of 5 subsets |
|---|---|---|
| All-at-once | the whole trajectory in one question | 52.0% (Table 1) |
| TRIAGE | ranks steps by suspicion and structural impact | 48.3% (Table 1) |
| FALAT | dependencies to separate origin, propagation and correction | 39.7% (Table 1) |
| BLADE | dependencies pick the handoffs; audits label how each source was used | 57.6% (Table 1) |
Why audit selectively instead of checking every pair? Take a trajectory of 20 events (illustrative length), and assume any event could depend on any earlier one. Auditing every earlier-to-later pair means 20 × 19 ÷ 2 = 190 handoff judgments. BLADE audits at most 2 sources for each of events 2 to 20, and only sources whose depends-on probability clears 0.3, so at most 19 × 2 = 38 judgments, which is at most 20% of the exhaustive count in this illustration. That matches what the paper measured: dependency-guided auditing examined 13–23% as many handoffs as exhaustive auditing and still matched or beat its step accuracy on three of the five subsets. Fewer audits did not cost accuracy because the selected handoffs are the ones where the error sits: on the Magentic-One subset, 70% of audited handoffs into the annotated decisive step were labeled breach, against 4% of those into earlier steps. The paper estimates the whole pipeline at about $1.9 in model cost for all 1,228 runs.
Removing parts of the method one at a time supports the central claim. Adding dependency context alone to the All-at-once question gave mixed results, while adding the audit findings raised step accuracy by 2.4–3.4 points on every subset tested. Shuffling the findings onto the wrong event pairs made accuracy worse, so the gain comes from evidence attached to the right handoff, not from a longer prompt. For your own agent team, the practical lesson is to log each handoff as its own span with the source it consumed, so that this question can be asked afterwards.
Goes deeper in: Agent Engineering → Agent Teams → Handoffs
Continue in trackAgent Teams: what gets lost when one agent hands work to anotherRelated explainers
- STRACE: causal trace localization — the dependency-graph approach this one argues is not enough on its own
- TELBench: span-level error localization — finding the first-error span inside a single agent's run
- Magenta: error-attribution routing — classifying why an attempt failed before deciding how to retry