What It Is
Commerce-1 is a small language model trained on Machine Commerce behavioral traces. It is built to predict and classify agent decision quality in e-commerce flows. The model runs as a lightweight inline evaluator inside agent pipelines, processing each action sequence as it happens and outputting a failure probability and failure mode classification before the failure reaches production.
Commerce-1 is not a general-purpose evaluation model. It is trained specifically on the behavioral signature of commerce agents operating in real e-commerce environments: product search, cart management, checkout flows, authentication, form completion, and order confirmation. The training distribution comes from Machine Commerce, our ongoing behavioral trace collection program across live commerce sites.
The Problem It Solves
Commerce agents fail in ways that are difficult to detect from server logs alone. A form-fill failure, an authentication friction event, and a product disambiguation error can all produce the same observable outcome: the session ends without a completed order. But they have distinct causes and require distinct fixes.
Identifying which failure mode occurred typically requires replaying the session, examining the agent’s action trace step by step, and making a judgment about where the decision quality broke down. At any volume, this is not practical. Pipelines that log thousands of agent sessions per day cannot manually inspect each one.
Commerce-1 closes this gap. It classifies the failure mode in real time, enabling targeted intervention rather than full session restart. A pipeline that knows it is looking at an authentication friction event can route the session differently than one that has hit a product disambiguation failure. Commerce-1 makes that distinction available at inference time, not retrospectively.
How It Works
Commerce-1 runs as an inline evaluator on each agent action sequence. It takes a partial session trace as input: the sequence of actions the agent has taken so far, the current state of the page or environment, and any available context about the target task. It outputs two values: a failure probability score between 0 and 1, and a failure mode classification drawn from a taxonomy built on the Machine Commerce trace dataset.
action
eval
threshold
intervene
The model is designed to run in-process with sub-50ms latency on standard hardware. It does not add meaningful overhead to agent pipelines. Because it is small and self-contained, it can run on the same infrastructure as the agent itself, without requiring a separate inference endpoint or external API call.
Commerce-1 is designed to be composed, not to replace the agent’s own decision logic. The typical integration is a conditional check after each significant action: if Commerce-1 scores the current trace above a configurable failure probability threshold, the pipeline pauses for review or routes to a fallback handler. Below the threshold, execution continues normally. The threshold is tunable per deployment depending on how the operator wants to balance latency against failure prevention.
on standard hardware
categories classified
covered end-to-end
required at runtime
Status
Commerce-1 is in active development. Training data collection is ongoing through Machine Commerce, which runs continuous behavioral trace collection across real e-commerce environments. The dataset covers the full range of commerce agent tasks and includes both successful and failed sessions with structured annotations describing the failure mode, failure point, and contributing factors.
The first model release is expected alongside CartBench in early 2027. CartBench is our evaluation benchmark for commerce agent pipelines, built from the same trace data. Commerce-1 and CartBench are being developed in parallel so that the model’s failure classification output is directly interpretable against the benchmark taxonomy from day one.
We are offering early access to teams building commerce agent pipelines who want to shape the training distribution, the failure mode taxonomy, or the integration API. Early access participants receive the model before general release and can work directly with us on fine-tuning for their specific pipeline and commerce environment.
Get Early Access
If you are building a commerce agent pipeline and want to be part of the early access program, contact us at founders@generalmachines.ai.
Include a brief description of your pipeline: what tasks your agent handles, the environments it operates in, and the failure modes you are currently trying to catch. We are particularly interested in teams working at volume, where manual session review is already a bottleneck.