Auto-building agents

Describe a goal. Super auto-builds it into AgentScript, a scripting language for reusable agent workflows. Build the workflow once, inspect it, and rerun its saved logic instead of asking a model to reinvent every step.

Scripted control flow makes repeated execution more predictable and reduces opportunities for model errors to accumulate across long tasks. For illustration, an independent 1% error probability per decision produces a 99.3% chance of at least one error over 500 decisions. This is a hypothetical example, not a measured hallucination rate or a guarantee that every error causes failure.

Use models where intelligence is needed: generating text, images or video, interpreting a page, and making judgments. Let code handle the reusable workflow around those calls. Replacing 100 comparable model calls with one means 100x fewer calls; actual savings depend on the workflow, tokens and providers. Script execution avoids per-step inference charges, but computing, Super access and remaining model calls can still cost money.

Build and run workflows with your connected Chrome extension or Mac app, inspect results, and continue projects across chats. Scripts and changing websites can still fail, so completion is verified from observed results. Requires a Super account and a connected client; Chrome execution requires Growth access or an authorized manual grant.