A single agent in a chat window is a demo. A team of agents you can actually supervise is a tool. The gap between the two is mostly architecture — where the work runs, and how you see it.

CodeSplash is built around the idea that agents are parallel workers, not a conversation. You dispatch several at once, each in its own session, and you inspect their output in real integrated apps — the editor, the terminal, the file tree, git — instead of scrolling a transcript.

Agents as workers, not a chat

The mental model is closer to a build queue than a chatbox. You describe tasks, hand them to agents, and let them run concurrently:

TypeScript
const tasks = [
  { id: "refactor", prompt: "Extract the auth flow into its own module" },
  { id: "tests", prompt: "Add coverage for the payment webhook" },
  { id: "docs", prompt: "Document the new config options" },
];
 
const sessions = await Promise.all(
  tasks.map((task) => agents.spawn(task)),
);

Each session is isolated, so one agent editing files never trips over another. When a session finishes, its changes are already visible in the apps you'd use to review them.

Bring your own model

CodeSplash doesn't resell inference. You connect the AI accounts you already pay for, and CodeSplash orchestrates them. That keeps two things true:

  • No extra subscription. You're not paying a platform markup on top of your model bill.
  • Your choice of CLI. Point sessions at whichever agent CLI you prefer — there are more than twenty to choose from — without rebuilding your workflow.

Why the integrated apps matter

The reason parallelism usually fails in AI tools isn't the models — it's supervision. If you can't see what three agents did, three agents is worse than one. CodeSplash pairs every session with the apps you'd normally use to check work:

  1. The editor shows the diff in context.
  2. The terminal sits next to the code it ran.
  3. Git and search let you verify across the whole repo.

Parallelism only pays off when review keeps up with it. That's the whole design.

What's next

I'll go deeper on session isolation and the review loop in future posts. If you want them as they land, subscribe below.