The Autonomous Loop

Most analytics tools let an AI agent read your data. A few let it create an object. Accelerate lets an agent run the entire optimization loop — propose a change, test it, read the live results, and ship the winner — over MCP, with no one in the editor. This page explains that loop, and the guardrails that keep it honest.

The four stages#

A test created through the Abilities API goes through the same lifecycle a human runs in the block editor — start to finish, through one set of abilities:

Because traffic follows the winner — Accelerate allocates with multi-armed bandits — the loop wastes less of your audience on losing variants while it runs, and returns a clean confirmatory reading before any winner is declared.

Design & Evolve#

The loop doesn't only test ideas you supply — with the Toolkit, it can write and improve them:

• Design composes a fresh, on-brand version of a block (or a whole new section) from a brief like "make me a bolder hero." It grounds every variant in your site's own design grammar — global styles, synced patterns, representative pages, and your real media — rather than a generic template, then offers to put the result in a test.
• Evolve runs the loop round after round on one block: each round fields bold new challengers, tests them, keeps the winner, and builds the next round on it — until the improvements stop. The number of variants per round scales with your traffic, so every arm still gathers enough data to be measured.

To read and back up the content it's about to change, the toolkit uses the read abilities — get-content returns raw block markup (including drafts), and get-variants returns each variant's markup — so a change is never reconstructed from memory.

Attended or autopilot#

The loop runs at two levels of autonomy:

• Attended (default). You confirm each round's challengers and each winner. Every change is surfaced before it ships.
• Autopilot. You grant a standing approval up front and the loop iterates unattended, reporting progress as it goes. Autopilot removes the per-step click, not the discipline: a winner is still only declared at the 95% probability-to-win threshold with enough conversions on every arm, and inconclusive rounds keep running rather than crowning a false positive.

Autopilot is the one place Accelerate acts without a per-change confirmation, so treat starting it as the approval it is — see Permissions → Autopilot is a standing approval.

Who made each test#

Experiments carry their origin, surfaced in the experiments overview and the block editor as up to two independent badges:

• AI — the test was created by an agent (shown permanently).
• ● Autopilot — an autopilot loop is actively managing this test right now (shown only while it runs). It's the signal not to interrupt a test the loop is mid-way through.

A human-created test carries no badge. Provenance (who made it) and activity (is the loop live) are separate, so they can appear together.

A worked example#

The whole loop, from one instruction:

You: "Set up an A/B test on my homepage hero — try a bolder, outcome-focused version — watch it, and ship the winner."

1. The agent calls get-content to read the current hero, then create-ab-test with a control and a challenger. The test starts running immediately.
2. Over the following days it polls get-experiment-results, reporting each variant's probability to win as the bandit shifts traffic toward the leader.
3. When the challenger sustains a 95% probability to win with enough conversions, the agent calls stop-experiment to declare it — and the winning hero is served to everyone.

Put the same block on autopilot and step 1 repeats for round after round, each new challenger built on the last winner.

Where to go next#