Salesforce AI Echoing at 70%: Metrics Miss It

Two LLM agents walk into a workflow. One is supposed to negotiate, the other is supposed to advise, and after enough back and forth they start doing the computational equivalent of swapping lanyards at a conference. Salesforce AI Research has a name for this: echoing, where agents abandon assigned roles and mirror their conversational partner. The unsettling part is not that agents fail, because welcome to software, please take a pager. The unsettling part is that the paper reports echoing rates as high as 70%, while Salesforce AI Research says standard task-completion metrics can miss the identity boundary failure entirely.

What Salesforce is bringing to ICLR 2026 According to Salesforce, its AI

Research team will present 21 accepted papers at ICLR 2026, which runs April 23 to April 27 at the Riocentro Convention and Event Center in Rio de Janeiro, Brazil. The company says its work on agent identity failures was accepted to the Agents in the Wild: Safety, Security, and Beyond workshop. That puts Echoing: Identity Failures when LLM Agents Talk to Each Other in the less glamorous, more useful corner of AI research: the place where agents stop being demo confetti and start being systems that need invariants. Salesforce AI Research’s publication page lists the paper under ICLR Workshop 2026 and names Sarath Shekkizhar, Romain Cosentino, Adam Earle, and Silvio Savarese as authors. The same page says the paper formalizes echoing as an emergent failure in agent-agent interactions where models abandon their identity and mirror conversational partners. That phrasing matters because this is not just bad roleplay. It is a boundary failure in a system whose whole premise depends on roles staying meaningfully distinct.

The awkward mirror test, according to arXiv According to the

arXiv abstract for Echoing: Identity Failures when LLM Agents Talk to Each Other, the authors studied autonomous agent-agent conversations and found echoing across major LLM providers. The paper reports echoing rates as high as 70%, depending on the model and domain. It also reports that echoing persists in advanced reasoning models at 32.8%, and that increased reasoning effort did not reduce those rates. Apparently, thinking harder is not the same as remembering who you are, which is also true at family weddings. The arXiv abstract says the failure becomes more likely as interactions grow longer, specifically at 7+ agent turns. That is a useful warning for anyone building multi-agent workflows with long deliberation loops, review chains, negotiation agents, or synthetic committee meetings from hell. A single agent can pass a unit test, then drift when paired with another model that has a conflicting conversational gravity. This is why agent-agent evaluation is not just single-agent evaluation with extra chairs.

Why standard metrics are too polite, per Salesforce AI Research Salesforce

AI Research’s publication page says echoing can remain undetected by standard task-completion metrics. That is the paper’s most builder-relevant punchline, because task success can become a very expensive participation trophy. If agent A completes the booking, agent B summarizes the case, and both quietly mutate their identities along the way, the metric may still smile and stamp approved. The system did the thing, but not necessarily as the system you designed. That finding lines up with a broader research thread around multi-agent LLM fragility. In Why Do Multi-Agent LLM Systems Fail?, the authors report an analysis of five popular multi-agent system frameworks across over 150 tasks, with six expert human annotators. They identify 14 unique failure modes grouped into specification and system design failures, inter-agent misalignment, and task verification and termination. Echoing fits neatly into the uncomfortable middle: agents can communicate, collaborate, and still blur into each other like a badly trained improv troupe.

What builders should watch, according to the Echoing paper The

Echoing arXiv abstract says the authors introduce a protocol-level mitigation using targeted structured response, reducing echoing to 9%. That is a practical clue, not a magic amulet. Structured responses can force agents to keep identity relevant in the communication protocol, instead of trusting a prompt header to behave like a legal contract. If your agent architecture relies on please remain in role as the main guardrail, congratulations, you have invented a Post-it note for a tornado. For builders, the takeaway is straightforward: evaluate identity preservation as its own property, especially in longer agent-agent conversations. Track whether agents maintain role, objective, and authority boundaries, not just whether the final answer looks plausible. Watch whether vendors and frameworks add AxA identity tests, structured response protocols, and task metrics that distinguish completion from role collapse. Multi-agent systems are not doomed, but they do need better mirrors, preferably ones that do not start talking back in the wrong voice.