Core concepts

These are the customer-facing terms used throughout this guide. Here they are with a shopping site’s review summary feature as the example.

Term	One-line definition	Review summary example
Feature	The customer-visible product unit whose performance you judge, identified by a Feature ID	review.summary
Node	A point inside a feature that calls an LLM. Each node gets its own Agent Config. A feature whose calls are not split has one node	The call that picks reviews, the call that writes the summary
Variant	One way of running a feature. Decides which Agent Config answers	The Origin variant / a candidate that switches only the summary call to Claude
Agent Config	One node’s set of model + prompt + parameters. Saved in Testing and picked by variants. Its contents never change once created	GPT, a long summary instruction, temperature 0.3
Origin	The setup your code sends today. Created automatically on the first call, and kept as the Origin variant on the feature and as an Origin Agent Config on each node	GPT + a long prompt, as your code uses now
Routing	The ratios that split real traffic across variants	Current 90% + new candidate 10%
Event	What a user who saw a response actually did	summary_copied, purchase
Success metric	The one number that defines success for a feature	Summary success rate, revenue per AI dollar

How the terms connect

Routingsplits the traffic Variantgenerates the response Eventwhat the user did Success metricoutcome as a number

This is the loop inside one feature (review summaries). New candidates enter it only after being screened in Testing.

Origin

Origin is created automatically from the setup your code sends, and it lives in two places: the feature and each node.

Kind	What it is
Origin variant	The baseline variant, one per feature. When your code changes the model or parameters, a new Origin variant takes over the previous Origin variant’s ratio, and the previous one stays so you can read its past results
Origin Agent Config	The model and parameters your code sends, per node. It holds no system prompt and uses the request’s system prompt as sent. A node that a variant does not change responds with this setup

Good to know

• Features and events register automatically on the first call or first occurrence.
• The same user keeps getting the same variant until you change the ratios. Their experience stays steady, and the comparison between “the people who got A” and “the people who got B” stays clean per person.
• Comparison, routing, and success metrics all operate per feature, because the model that suits review summaries and the one that suits support replies are different.

Events and success metrics

The SDK installed in your site or app sends events in one line. A sent event connects automatically to the AI responses that user received.

abto.capture('summary_copied');
abto.capture('purchase', { value: 32000 });

You define what counts as success yourself, out of events and call records, and there are two ways to define a success metric.

Form	What it measures	Example
Ratio	The share of people using this feature who did the target behavior	Summary success rate = purchase 70% + summary copied 30%
Value	One number defined by a formula	Revenue per AI dollar = sum of purchase value ÷ sum of call cost

Designate one metric per feature as Primary and it becomes the Overview summary number and the default comparison basis on Compare.

See Dashboard at a glance for the customer tasks these terms support.