ASTRA - Agentic System for Ticket Resolution and Analysis

Enterprise AI agents are increasingly deployed as harnesses around frozen foundation models, where adaptation happens not through weight updates but through accumulated memory, retrieval over organizational experience, and iterative refinement from internal and external signals. In this context, we report on ASTRA, a multi-agent system for telecom fault analysis deployed over a corpus of approximately 130K historical tickets, and use it to study what such a system actually learns. In ASTRA, a central orchestrator coordinates three specialist agents, a TicketSimilarityAgent (dense retrieval + LLM reranking over historical tickets), a LogAgent (deterministic filtering + constrained LLM analysis of raw logs), and a DomainKnowledgeAgent (documentation retrieval via MCP), each producing bounded, verifiable artifacts. Their outputs are assembled into a claim-evidence intermediate representation that links every diagnostic claim to a verbatim source passage and prevents cross-attribution. Then a judge-orchestrator refinement loop iteratively improves coverage over bounded rounds by continually adapting to new feedback and contexts. Evaluated on 987 real-world telecom fault tickets across 7 product lines, ASTRA achieves a mean quality score of 4.13/5.0, with 59.9\% of reports identifying the correct fault area, and fabricated technical details appearing in under 3\% of low-accuracy reports (Accuracy $\leq$2). Stratification reveals that hardware-related root causes remain substantially harder than software or configuration faults (Cohen's $d$ = 0.80), exposing a fundamental limitation of text-based evidence for hardware fault diagnosis.