Humans& released Ampersand v0.1 through its Agents API for developers building tools and apps that foster collaboration. The company describes training in multi-user simulations where goals and preferences evolve over long periods. It says its evaluation measures progress toward users’ long-term goals minus interaction costs that conflict with their preferences. These are the observable design choices behind the release: the model’s target is collective and longitudinal, and its score includes the cost of how it intervenes.
Mechanism: The model is trained and evaluated in simulated multi-user settings where goals evolve; reported value combines progress toward long-term goals with penalties for interactions that conflict with preferences. The company embeds the model in simulated groups with seeded goals that can change over episodes of up to months of simulated time, then scores value as goal progress less accumulated interaction costs. That reframes model behavior as coordination under shifting context rather than a single-turn answer or isolated task result. Humans& reports Ampersand at 59.1% on Collaborative Intelligence at high effort and 39.0% cooperative success on CooperBench; it also acknowledges higher interaction costs than other leading models, including unwanted coaching.
Why now: The release makes this training and evaluation approach available to developers through an API, so it can now be tested in collaborative applications. This turns the collaboration claim into a product that developers can test.
First implication: Teams should evaluate collective progress and intervention costs together, rather than equating proactivity with collaboration. Teams evaluating collaborative agents should measure whether users and groups make progress together while tracking the cost of unsolicited intervention, rather than treating more proactive behavior as an unqualified gain.
Evidence boundary: All benchmark results and comparisons are reported by Humans&; the supplied source does not establish independent replication or real-world deployment outcomes. The supplied source does not provide independent replication or evidence of outcomes in deployed customer products.
First watch condition: Watch for independent evaluations or deployment results that test whether collaboration gains persist while reducing unwanted coaching across varied groups and tasks. Such evidence would show whether the benchmark framing generalizes beyond the company’s own evaluation.