Sarath Shekkizhar, Prathyusha Jwalapuram, Zixuan Ke, Vaidehi Patil, Jiayu Wang, Srijan Bansal, Trong Huy Nguyen, Dingfeng Jiang, Dongqi Liu, Ruochen Zhao, Wei Lu, Giuseppe Carenini +2
20 pages, 7 figures, 4 tables. Run logs: https://huggingface.co/datasets/a2a-salesforce/a2-failures
In agent-to-agent systems where agents serve different principals with private state, failures cluster into 20 identifiable modes, and three regularities hold across harnesses and models: agents only contact peers when prompts name them, agents never verify peer claims unprompted, and outcome-only scoring hides coordination breakdowns.
Over 2,500 recorded runs across 488 scenarios in eight families (legal, CRM, risk, software, security, commerce, negotiation, evolving enterprise), with agents in isolated Docker containers communicating via four MCP tools, varied one dimension at a time across harnesses (Claude Code, Codex, Cursor) and models (Opus-5, Sonnet-5, GPT-5.6 variants, Grok-4.6), scored at three layers (environment state, artifact correctness, principal-aligned conduct), with taxonomy derived via grounded-theory coding and logs released.
Qualitative and descriptive rather than quantitative: three cross-cutting observations reported as holding across all configurations, plus scenario counts per family (e.g., 224 legal, 83 evolving-enterprise) and scaling effects such as filing-before-reply worsening with team size; no aggregate failure rates, baseline comparisons, or effect magnitudes are given in the provided text.
Authors concede the taxonomy is not exhaustive, scenarios are a diagnostic instrument rather than a benchmark, and causality is limited to localizing failures to system layers rather than broad generalization; they also note observability gaps (L0 log gaps can falsely trigger fabrication flags). Unstated: authors are Salesforce employees evaluating their own A2A framework, scenarios are self-designed, and no other taxonomy's inter-rater reliability or coding reproducibility is reported in the provided text.
A skeptic would note this is a company-built diagnostic suite where the same team designed the scenarios, ran the agents, and coded the failures, with automated tagging doing much of the classification; the headline findings (agents do what prompts say, don't verify) could partly be artifacts of harness prompting conventions and scenario design rather than fundamental A2A properties, and no success-rate or intervention-ablation data is shown to prove the taxonomy predicts anything.
Little or nothing yet for economics in the next 12 to 24 months: this is an early diagnostic of an interaction pattern (independent agents negotiating across organizational boundaries) that is not yet a commercial deployment pattern; its near-term value is as a warning that agentic enterprise ROI claims premised on multi-agent coordination lack evidence, and that outcome-only vendor benchmarks would mask these failures.
Technical Capability: Shows that multi-agent reliability failures are structural to cross-boundary interaction (discovery, verification, reconciliation), not just model-capability gaps, meaning better base models alone will not fix agentic coordination.
Labor & Knowledge Work: If enterprise agent deployments silently fail at coordination and verification layers while aggregate scores look fine, the gap between agentic-productivity marketing and delivered work will be wider than buyers realize.
Regulatory & Legal: The conduct layer (unauthorized disclosure, misrepresentation undetectable without environment logs) gives regulators and auditors a concrete list of what cross-organizational agent deployments would need to log and verify.