humans& Trains Ampersand to Collaborate and Hits 92.2% on Hard Coding Benchmarks

In a benchmark report published by humans&, multiple Ampersand instances working together matched the top reported score on a difficult repository-level coding test. The model received collaboration-focused post-training with no coding-specific objective reported, suggesting that coordination skills can transfer to software engineering tasks.

The multi-agent setup also reduced median completion time under matched serving conditions, despite making substantially more model calls. That speed comes with higher inference costs. All figures were produced by humans&, and the report includes no independent replication.

Collaborators close the coding gap

SWE-bench Pro v2 HARD-51 contains 51 difficult software issues that agents must resolve in real repositories. DeepSWE v1.1 provides a separate 113-task evaluation. The mini-swe-agent entries use an external harness that repeatedly lets the model inspect, edit, and test a repository.

On HARD-51, the multi-agent result tied Fable 5.1 and Gemini 4 Argon in the lab’s comparison while exceeding Astra, Muse Wonder, and Kimi K3. Adding collaborators improved Ampersand’s score by 11.8 percentage points over its single-agent Agents API run.

On DeepSWE at the same high-effort setting, collaboration added 9.8 percentage points and raised average task cost from $4.19 to $14.03. Compared with the max-effort solo run, the gain was 5.3 points at more than twice the average cost. The effort labels refer to API presets that allow different amounts of inference work.

More calls, less waiting

Under matched serving conditions, a four-agent team completed tasks sooner while making roughly 2.5 to 3 times as many model calls. Parallel repository exploration and editing reduced wall-clock time even as total model activity increased.

The latency result reflects concurrency rather than reduced computation. Multiple agents can investigate separate files or hypotheses at once, then coordinate their changes. The DeepSWE costs show the resulting tradeoff directly; the report does not provide an average multi-agent cost for HARD-51.

Collaboration training transfers

humans& post-trained Ampersand from Z.ai’s GLM 5.3 using long-horizon, multi-agent reinforcement learning. Training rollouts placed the model in groups with shifting goals, incomplete specifications, and disagreement. The available material describes the post-training target but does not establish whether GLM 5.3 encountered code during pretraining.

The training stack combines three mechanisms intended to make those long group interactions workable:

• Persimmon user simulator. A trained user model generated participants whose goals and opinions could change. Those simulated users rejected unhelpful or ingratiating responses, forcing Ampersand to negotiate and coordinate through disagreement.
• CORO reinforcement-learning loss. Contrastive Off-policy Reward Optimization compares trajectories with one another, removing the need to calibrate a separate value estimate. Off-policy rollouts may come from earlier model versions, so this approach targets instability caused by stale training data.
• Dual-ascent regularization. The system adjusts its regularization coefficient as a Lagrange multiplier, keeping divergence between the policy being trained and the policy that generated the samples within a fixed budget.

CooperBench exposes the coordination gain

CooperBench evaluates agents implementing potentially conflicting features in the same repository. In humans&’s runs, GLM 5.3’s success rate fell from 52.5% on solo tasks to 17.3% on cooperative tasks. Ampersand reached 39.0% cooperative success, and the lab reports that its agents communicated throughout each task to divide work and reconcile changes.

A solo Ampersand score is absent from that comparison, preventing a within-model estimate of CooperBench’s collaboration gain. The reported result still supplies a plausible mechanism for the coding scores: agents that exchange findings and coordinate patches are less likely to duplicate work or overwrite one another’s changes.

Calling Co-Coding through the API

Access to multi-agent mode runs exclusively through the OpenAI-compatible Agents API . Developers can configure the official OpenAI SDK with Ampersand’s base URL and the ampersand-v0.1 model identifier:

After installing the OpenAI SDK, the same values can initialize a client:

The session abstraction stores conversation history on the server, allowing follow-up requests to send only new messages. humans& advertised a $250 trial credit for one week after the announcement. Current prices, rate limits, and session calls are listed in the API reference .

Where the evidence stops

The benchmark package combines Ampersand’s base coding ability, collaboration post-training, agent harness, effort setting, and serving configuration. It therefore cannot isolate the contribution of each component. The comparisons also lack independent replication, confidence intervals, and complete cost data for the named competing systems.

Ampersand’s lower solo scores indicate that coordination drives much of the reported gain. For repository tasks where resolution rate and elapsed time carry more weight than inference cost, the multi-agent configuration offers a credible option. Broader evaluations across code review, testing, security work, and long-running production changes would show how well that advantage extends beyond these benchmark sets.