STAR: A Stage-attributed Triage and Repair framework for RCA Agents in Microservices¶

🕒 Published (v1): 2026-05-15 03:44 UTC · Source: Arxiv · link

Why this paper was selected

STAR: stage-attributed triage and repair improves reliability of RCA agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

STAR is a meta-debugging layer for LLM-based root cause analysis (RCA) agents in microservice AIOps that treats agent failure as a stage-localizable reasoning bug rather than an opaque end-to-end error. It audits a structured four-stage RCA trace, identifies the earliest decisive faulty stage via counterfactual replay, patches only that stage, and replays all downstream stages to eliminate error propagation.

Problem¶

LLM-based RCA agents for microservices are brittle: errors introduced at evidence collection, hypothesis formulation, or causal analysis cascade through subsequent reasoning steps, yet fine-grained step-by-step debugging is both unreliable (traces are noisy) and expensive (full regeneration costs). No existing framework explicitly models where in a structured agent workflow the failure originates so that repair can be targeted and minimal.

Method¶

STAR wraps any existing RCA agent workflow (implemented on LangGraph) and operates over a structured four-stage trace \(A = (\text{EP}, \text{HS}, \text{AS}, \text{DR})\):

Stage-wise Audit: A weighted scoring function \(S = \sum_k w_k s_k\) evaluates each stage artifact against RCA-specific constraints (evidence time-window coverage, hypothesis grounding, topology-feasible causal paths, confidence calibration) and produces per-stage diagnostics \(d = \{d_{S_1}, \ldots, d_{S_4}\}\).
Fast/Slow Routing: Score \(S\) is compared against thresholds \(\tau\) and \(\tau - \epsilon\). Pass (\(S \ge \tau\)): accept original output. Fast Path (\(\tau - \epsilon \le S < \tau\)): single constrained local patch on the highest-severity stage. Slow Path (\(S < \tau - \epsilon\)): full replay-validated localization.
Decisive Stage Localization (Slow Path): For each stage \(s\), a stage critic \(C_s\) proposes \(K\) structured patch candidates \(\{P_s^{(k)}\}_{k=1}^K\). Each candidate is evaluated via counterfactual replay: \(\Delta S(s,k) = S(\text{Replay}(A, s, P_s^{(k)})) - S(A)\). STAR selects \(s^* = \min\{s \mid \max_k \Delta S(s,k) \ge \delta\}\)—the earliest stage whose patch yields measurable downstream improvement. An INSUFFICIENT_EVIDENCE_EXCEPTION forces rollback to \(S_1\) if no topology-feasible causal chain can be constructed.
Patch-and-Replay Repair: The best patch at \(s^*\) is committed and \(\text{Run}_{(s^*+1) \to 4}\) re-executes all downstream stages conditioned on the corrected upstream artifact. Stage-specific patch operators are defined in Table I (e.g., for EP: time-window expansion, modality re-query; for AS: topology-reachable chain reconstruction).
Self-Evolving Repair Memory: Successful repairs \(\langle q_i, s_i, P_i, \Delta S_i \rangle\) are stored and retrieved to prime stage critics with historically effective patch templates.

Key Contributions¶

STAR framework: stage-attributed audit + decisive-stage localization + patch-and-replay for LLM-based RCA agents, built on LangGraph.
Fast/Slow Routing: budget-aware triage that avoids expensive replay on near-miss traces.
Counterfactual candidate evaluation: principled criterion (\(s^*\) minimality + replay improvement \(\ge \delta\)) for identifying the causal root of agent failure rather than a merely suspicious stage.
Self-Evolving Repair Memory that accumulates incident-pattern-to-patch mappings to accelerate future repairs.
Evaluation across two agent workflows, three foundation models, one public AIOps benchmark (HipsterShop2, 15 fault types, 40 pods), and one real-world production dataset (12-service power-enterprise platform).

Results¶

The paper text is truncated before reporting quantitative tables. From the abstract and experimental framing, the reported findings are: - STAR consistently improves Acc@1, Acc@3, Acc@5 (root cause localization) and fault type classification over base RCA agent baselines on both Dataset A and Dataset B. - STAR identifies the decisive faulty stage \(s^*\) with high precision (specific precision values not present in provided text). - Most initially incorrect traces are repaired within one or two replay iterations. - Ablation confirms that both Fast/Slow Routing and counterfactual decisive-stage localization contribute independently to repair efficiency and final accuracy.

Limitations¶

Evaluation uses only two RCA agent workflow designs; generalizability to arbitrary multi-agent topologies is unverified.
Stage critic prompts and patch operators are hand-designed around four fixed artifact types (EP/HS/AS/DR); adapting STAR to workflows with different stage decompositions requires redesign.
The audit score weights \(w_k\) are not automatically tuned; miscalibrated weights can mis-route traces to Fast or Slow paths.
Self-Evolving Repair Memory is not evaluated standalone; its contribution relative to fresh counterfactual search is unclear.
Full numerical results are not available in the provided text, limiting external verification.

Relevance to Harnesses / Meta-Harnesses¶

STAR is a canonical example of a meta-harness: it does not perform RCA itself but wraps an existing agent workflow to audit, localize, and repair its intermediate outputs without rerunning the full pipeline from scratch. Its stage-structured trace representation (\(A\) as a typed, replayable program state) and LangGraph-backed node-level replay are direct harness design patterns—analogous to how test harnesses intercept and verify intermediate program states. The Fast/Slow Routing and budget-aware repair reflect a harness design concern—controlling orchestration cost—that is absent from the base agent. For researchers tracking meta-harnesses, STAR illustrates how decomposing agent execution into typed artifacts with explicit schema contracts enables systematic, programmatic repair loops over LLM-generated reasoning.