Skip to main content
This takes about ten minutes and needs no terminal. You end with two agents’ work judged side by side.
1

Sign in

Open tracks.alpic.ai and click Sign in.Your first sign-in creates your organization with $10 of free credit. If a teammate invited you, you join their organization instead: see organization and sharing.
2

Check your balance

Open your profile menu at the bottom of the sidebar, then Billing. Your balance is what’s left to spend on runs.To add more, click Add funds, pick an amount and pay by card. See billing.
3

Say what to test

Open Experiments. Your first visit asks What would you like to test? Pick one way in:
  • A template: click one of the questions after “Can agents”, such as find pricing or get started, enter your website’s domain and click Confirm.
  • Your own words or a document: describe the task, or paste or upload a test plan, a spec or a checklist, then click Continue. A model turns each task it describes into a scenario, with evaluations in your document’s words.
  • Start from scratch: an empty scenario you fill in yourself.
Later, New experiment offers From scratch and From a document.
4

Review the experiment

The editor walks you through four steps. Changes save on their own.
  • Scenarios: each one is a prompt for the agent and its evaluations, the yes-or-no questions that decide whether a run passed. See write evaluations.
  • Agents: Claude Code and Codex are already listed, each with its default model. Add or change agents here.
  • Variants: each set of agent tools the agents get, picked from your library. Every experiment starts with one set without tools, as the baseline. Add a set with your MCP server or skill to compare with and without it: see measure your MCP server.
  • Run settings: how many times each run repeats, and each run’s time limit and budget. By default, 1 repetition, 10 minutes and $1: one scenario with two agents makes 2 runs, spending up to $2.
If something is missing, Launch takes you to it.
5

Launch

Click Launch. The dialog shows how many runs start and the most they can spend, next to your balance. Click Launch again to start them.
6

Read the results

You land on the execution page. While runs work and get judged, it fills in on its own:
  • Pass rate: the share of judged runs that passed.
  • Runs and Total spend.
  • Results: the pass rate for each agent, and for each scenario and set of agent tools when you have several.
  • Agent tool use: how many runs given each skill, MCP server or command-line tool used it.
Click a cell to list its runs, then open one to read its conversation, its evaluations and the tools it called.
You now have two agents compared on the same task. See reading results for what each number means.

​
Next steps