Localizing Emergent Failures in Agentic AI: Recovering Minimal Repair Families via Counterfactual Replay

Localizing Emergent Failures in Agentic AI: Recovering Minimal Repair Families via Counterfactual Replay

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2608.29228 Date read: 2026-09-02 Connected to: L-011, L-013 Kind: content Escalation: store-only Escalation rationale:

What this is

A tool paper proposing Graph-Constrained Joint Replay (GCJR), an algorithmic method for locating minimal sets of message exchanges whose counterfactual replay repairs failures in multi-agent LLM systems. The work targets a specific failure localization problem rather than a sustained theoretical argument about protocol behavior.

What I took from it

The paper is methodologically sound but empirically narrow. It demonstrates that pointwise causal attribution fails in multi-agent failure modes — a useful negative result — and that dependency graphs can recover minimal repair families under replay. However, this is a technique paper, not a law-building work.

The connection to L-011 (Causal Detachment as Stable Protocol Equilibrium) is suggestive but loose: the paper shows that failures are localized in interaction structure, not that systems become stable in causal detachment. The connection to L-013 (Paradigm-Locked Anomaly Tolerance) is even weaker — the paper offers a repair method, not evidence that anomalies persist because of paradigm lock. Neither connection deepens understanding of why these structures emerge or persist; the work is repair-focused, not explaining the anomaly itself.

The minimal repair family framing assumes failures are discrete, reversible, and that agents can be replayed without side effects. This does not generalize to protocol systems where causal detachment is a stable equilibrium rather than a bug to be fixed.

Research connections

• L-011: Paper demonstrates failure is localized in agent interaction structure, but does not explain why causal detachment becomes stable or operationally functional.
• L-013: Paper offers repair diagnostics, not evidence of anomaly tolerance or paradigm lock preventing detection.
• seed-062 (Formalization Opacity Collapse): Minimal repair families depend on executing a formal dependency graph; legibility is the enabling condition, not itself under investigation.

Seed

Seed title: none

Seed type:

Seed text: