Agent recovery depends on observable anomalies, verifiable task constraints, and evidence of root causes. Inference systems also face hardware partitioning and numerical consistency trade-offs.
Combining telemetry with deterministic checks
Real-Time Detection and Repair of LLM Agent Failures · Sunny Dubey · 2026-08-03
Agents can loop, cascade tool errors, or silently omit required work. Sunny Dubey trains one-class echo-state networks on healthy trajectories, combining step text and action telemetry with CUSUM alarms. A hybrid adds Mahalanobis detection, and a content-grounding channel checks tool results without invoking a second LLM at every step.
Across 2,823 episodes in three frameworks and several models, pooled detection of content corruption rose from 0.28 to 0.59, but only when corruption was visible in the telemetry. A separate deterministic path checks required calls and whether reported totals follow from tool results. On the same experimental set, the numerical check caught 60% of failures, rising to 96% with the coverage check, with zero false positives among 63 healthy examples. Rolling flagged runs back to the last fact-gathering step and rerunning recovered 45% of failures versus 16% for resampling. These figures apply to the paper’s particular tasks and conditions.
Organic failures were harder: in 30 high-temperature episodes the monitor caught one of three fabrications and falsely alerted on 36% of healthy runs; a completion check caught omissions. Cross-model transfer requires recalibration, while slow goal drift and plausible but incorrect external values remain blind spots. The operational lesson is to combine cheap anomaly signals, task constraints, and external checks rather than treat an alarm score as proof of correctness.
CUADebug: locating root causes in interface trajectories
CUADebug: Diagnosing and Repairing Computer-Use Agent Failures · Weijia Zhang 等 · 2026-07-31
A final screenshot may show failure while its cause lies several steps earlier. Weijia Zhang and colleagues propose CUADebugger, which selectively inspects before-and-after screenshots and actions, then returns an error subtype, root step, evidence, and correction for re-execution. Their human-annotated OSWorld set contains 204 failed trajectories, including 110 attributed to task reasoning and control.
On the main Claude trajectory split, joint subtype-and-step accuracy with Gemini 2.5 Pro rose from 11.2% to 19.6%. In single re-execution, machine RCA achieved 28.47% success versus 13.89% for history-only continuation. Restart locations and prompt formats differ between conditions, so the difference cannot be attributed entirely to diagnosis. The strongest evidence concerns OSWorld and Claude trajectories; other agent splits are smaller. Unlike telemetry monitoring, this work uses visual and action evidence to explain an interface failure after it occurs.
FailFast: stop early while retaining useful edits
Fail-Fast, Restart-Smart: Early Failure Prediction and Restart for SWE Agentic Tasks · Chenyu Wang 等 · 2026-08-04
Chenyu Wang and colleagues use a 0.6B monitor to predict failure from coding-trajectory prefixes. A fresh run of the same model receives the previous edits as an optional, initially disabled overlay that it can inspect, apply, or revert, rather than inheriting the failed reasoning. On SWE-bench Verified, stopping early saved up to 20.4% of execution tokens at a 5% false-positive budget. For Qwen3.6-27B, restarting at a 25% false-positive budget raised resolution by 5.2 percentage points with 43.8% more net tokens. The evaluation uses mini-swe-agent and SWE-bench Verified. Its stop-and-recovery trade-off complements CUADebug’s visual root-cause localization.
How MoE reduction order changes later behavior
From Expert Reduction to Behavioral Divergence: Tracing Numerical State through Sparse MoE Inference · Tianyang Zhu · 2026-07-30
Tianyang Zhu freezes local DeepSeek-V4-Flash state and changes only finite-precision expert reduction, tracing differences through hidden states, routing, and persistent state across tokens. At one layer fork, 720 A-mode orders produced 10 continuations; 360 structural classes under another mode produced 11. Exact state reconstruction checks the propagation chain for one controlled branch, showing that identical emitted tokens need not imply identical persistent state. The study covers one checkpoint, native runtime, and finite states; it measures neither production incidence nor the overhead of the stable scheme. It offers a causal case for numerical conformance testing, not an estimate of operational failure rates.
HeteroPanacea: when inference disaggregation pays
When Does Disaggregation Pay? Simulating Prefill–Decode–Attention–FFN Specialization for Agentic LLM Inference · Przemyslaw Forys 等 · 2026-08-04
Przemyslaw Forys and colleagues use an event-driven simulator to compare colocated serving, prefill/decode separation, and four-way attention/FFN separation while searching parallelism, precision, and stage-specific hardware. They simulate up to 75% greater throughput for a current-GPU scenario and up to 2.06× with custom NPUs. Disaggregation reaches parity only for sufficiently prefill-heavy workloads, and the four-way split beats two-way splitting only with enough hardware specialization. Components were validated on an 8×B200 node, but end-to-end scheduling and batching were not validated, and the GPU economics depend on a particular price offer.