📋 Executive Summary
• Production Agent Failures Get a Taxonomy — First longitudinal study of 22 real-world production agent incidents reveals "fail-plausible" LLM failures that fabricate believable narratives to hide errors. 70% caught only by human observation, not tests.
• Agent Harness Design Becomes Systematic — HarnessX introduces a composable harness foundry with formal substitution algebra + RL-driven evolution. +14.5% average gain across 5 benchmarks without model scaling.
• SLO-Driven Model Selection Enters Production AI — PLAIground achieves 91.3% accuracy under cost/latency SLOs where fixed-model baselines overshoot budgets by 21×. Directly applicable to banking compound AI deployments.
🔬 Deep Dives
📄 When Errors Become Narratives: A Longitudinal Taxonomy of Silent Failures in a Production LLM Agent Runtime
Wei Wu | arXiv: 2606.14589 (cs.SE) | June 12, 2026
METHOD: Eight-week longitudinal study of a production LLM agent system running ~40 scheduled jobs across 8 LLM providers with 4,286 unit tests and 827 governance checks. 22 incident postmortems with full root-cause analysis yielded a 5-class taxonomy of silent failures.
KEY RESULT: The most dangerous failure class (Class D: "Chained Hallucination and Fabrication") is unique to LLM systems: the agent not only fails but transforms the failure into a fluent, plausible narrative delivered to the user — termed "fail-plausible." About 70% of all silent failures were caught by human user-view observation — not tests, not audits, not governance checks. Incident latency ranged from 13 hours to 60 days. Full defense framework, 22 public postmortems, and governance engine released on PyPI.
🎓 SO WHAT FOR ANDY: This is the paper APRA CPS 230 has been waiting for. When a Big 4 bank deploys agent systems into production, "fail-plausible" is the nightmare scenario: an agent fabricates a regulatory filing, masks a transaction error, or creates plausible-but-wrong risk analysis — and no automated test catches it. The paper's defense framework provides a ready audit framework. Immediate recommendation: every production agent deployment needs a dedicated "user-view monitor."
Novelty: 5/5 | Actionability: 5/5 | 19/20
📄 HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry
Tingyang Chen et al. | arXiv: 2606.14249 (cs.AI) | June 12, 2026
METHOD: HarnessX treats agent runtime harnesses (prompts, tools, memory, control flow) as composable artifacts with a formal substitution algebra. Typed harness primitives are assembled and evolved through AEGIS, a trace-driven multi-agent evolution engine bridging symbolic adaptation and RL. Trajectories feed back into both harness updates AND model training signals.
KEY RESULT: Across ALFWorld, GAIA, WebShop, τ³-Bench, and SWE-bench Verified, HarnessX yields an average gain of +14.5% (up to +44.0%). Gains are largest where baselines are lowest. Agent progress need not come from model scaling alone. Codebase to be open-sourced.
🎓 SO WHAT FOR ANDY: For banking: instead of hand-crafting agent prompts for each regulatory use case, HarnessX's substitution algebra lets you compose verified harness primitives and evolve them from production traces. The AEGIS engine's operational mirror between symbolic rules and RL directly maps to banking governance: symbolic rules encode regulatory constraints, RL optimizes within them.
Novelty: 5/5 | Actionability: 5/5 | 18/20
📄 PLAIground: SLO-Driven Runtime Model Selection for Compound AI Systems
Milos Gravara et al. | arXiv: 2606.14356 (cs.DC) | June 12, 2026
METHOD: PLAIground introduces the Compoundable AI Model (CAIM) abstraction — decouples task semantics from model implementations via Task and Data Contracts, enabling model switching without workflow changes. The Pixie algorithm dynamically selects the most suitable model for each task at runtime to satisfy SLOs on accuracy, latency, and cost.
KEY RESULT: Pixie achieves up to 91.3% accuracy while maintaining SLO compliance. Fixed-model strategies either overshoot cost/latency budgets by up to 21× or miss accuracy targets by 4%. The CAIM abstraction means models can be hot-swapped during execution without workflow redesign.
🎓 SO WHAT FOR ANDY: This is Compound AI Operations for banking. A fraud detection pipeline chains multiple AI tasks with different accuracy/latency/cost trade-offs. PLAIground's CAIM contracts formalize the interface so models can be swapped under SLO pressure. Directly relevant to APRA CPS 230 operational resilience: dynamic degradation while maintaining minimum accuracy SLOs.
Novelty: 4/5 | Actionability: 5/5 | 18/20
🔍 Emerging Themes
• Production Agent Reliability Engineering (NEW arc #12): Silent Failures Taxonomy + PLAIground + RedAct form the kernel of a new discipline: production engineering for agent systems. The shift from "can the agent do the task" to "will the agent fail silently, and can we detect it under SLOs" mirrors the DevOps → SRE transition of 2015.
• Agent Infrastructure Tooling Maturation (arc #7): HarnessX + AgentSpec extend the infrastructure theme with rigor: formal substitution algebra and typed component composition. After 6 weeks of infrastructure paper accumulation, we're seeing convergence toward systematic design principles.
• KV-Cache as Computation Primitive (arc #8): Parallel-Synthesis demonstrates direct KV-cache synthesis across parallel agent branches — 2.5–11× TTFT reduction. The arc has evolved from communication to marketplace to direct computation.
• ICML 2026 Papers Breaking: Two ICML 2026 acceptances in today's batch (MRAgent, PoLar). Conference acceptances increase confidence — these have passed peer review.
🌍 AI-Adjacent Breakthroughs
• RATS! Emergent Object Parts in Vision Transformers (2606.14701) — Register Attention Transformers spontaneously discover compositional object structure — heads, wings, talons — without any part supervision. +12 mIoU across 5 segmentation benchmarks. Relevant to explainable AI for banking.
• El Niño Could Be Strongest This Century (Science, June 11) — Growing odds of a powerful climate event. Direct implications for commodity prices, agricultural supply chains, and insurance risk models.
⚡ Quick Hits
• Parallel-Synthesis (2606.14672) — Direct KV-cache synthesis across parallel agent branches. 2.5–11× faster time-to-first-token. First proof that cache-level merge beats text concatenation. [17/20]
• AgentSpec (2606.14674) — Typed, composable agent scaffold framework. Key finding: agent performance is governed by scaffold compatibility effects, not isolated module strength. [16/20]
• LLM Agents Can See Code Repositories (2606.14061, ASE 2026) — Hybrid text+vision reduces token consumption 26% while maintaining accuracy. Vision is navigation, not replacement. [16/20]
• Memory is Reconstructed, Not Retrieved (2606.06036, ICML 2026) — MRAgent replaces static retrieve-then-reason with active Cue-Tag-Content graph reconstruction. +23% with lower token cost. [16/20]
• Skip a Layer or Loop It? (2606.06574, ICML 2026) — Training-free dynamic program-of-layers: pretrained LLM layers can be skipped or looped per input with BETTER accuracy using FEWER layers. [16/20]
Sources: arXiv API (cs.AI, cs.LG, cs.CL, cs.CV, cs.MA, cs.OS, cs.DC, cs.AR, q-bio.NC, quant-ph — 12th triple success in 14 sessions), HF Daily Papers (50 papers), Science.org.
23 papers deduplicated from prior sessions (Jun 11–14).