Trust Between AI Agents: Measuring Formation, Breakage, and Recovery, with Implications for Governing Multi-Agent Systems

A new study proposes a behavioral measure to quantify trust between AI agents, finding that several frontier models substantially reduce costly verification of teammates they deem reliable, with implications for governing multi-agent systems [1]. The research, submitted in 2026, introduces a framework based on a cooperative survival game where checking a teammate's work consumes resources but trusting an incorrect answer can be fatal [1]. Relative to a memoryless version of the same model, reduced verification provides an observable measure of trust [2]. When paired with a consistently reliable teammate, four of six frontier model snapshots — Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.1, and Gemini 3.1 Pro — reduced verification by roughly 60-85%, while two smaller snapshots showed little or no such adjustment [1][2]. Failures reversed this discount, but models diverged in their responses. Some concentrated renewed scrutiny on the teammate that erred, while others became more cautious toward the entire team [1]. Recovery from broken trust was slower than its initial formation, and clustered failures sustained suspicion far longer than the same number of failures spread apart [2]. The study found that models which formed trust verified less, decided more quickly, and achieved higher payoffs, while persistent over-verification was associated with indecision rather than safety [1]. The findings arrive as multi-agent AI systems draw increasing research attention. A separate 2026 study evaluating 12 multi-agent LLM collaboration topologies for software architecture design found that structural adversarial variants and cross-model review configurations ranked highest across 520 experimental runs, while parallel-merge approaches consistently underperformed due to token starvation and incoherent outputs [3]. That work used Claude Opus 4.6 and Claude Sonnet 4.6 among its three independent automated evaluators [3]. Claude Opus 4.6 and Claude Sonnet 4.6 are developed by Anthropic, a San Francisco-based AI company founded in 2021 by former OpenAI members with a focus on AI safety [7]. Anthropic's Claude series, first released as a chatbot in March 2023, is trained using "constitutional AI" techniques aimed at improving ethical and legal compliance [5]. Since the Claude 3 generation, models have typically been released in three sizes: Haiku, Sonnet, and Opus [5]. The company was privately valued at an estimated $965 billion in May 2026 [7]. A separate benchmark study of MRI physics knowledge found that Claude Opus 4.6 and Claude Sonnet 4.6 achieved overall multiple-choice accuracy between 93.2% and 97.1%, but accuracy fell sharply when answer options were removed — dropping to 58.4% to 61.1% for frontier models in stem-only conditions, and as low as 13.8% to 29.8% for vendor-specific scanner operations questions [4]. The researchers concluded that high multiple-choice performance can mask weak free-text recall [4]. The trust-measurement study's authors argue that trust dispositions can be measured before deployment and that calibration, rather than maximal suspicion, should be the central concern in governing multi-agent AI systems [1][2].