tavya

Run many agents in parallel

Run multiple Claude Code, Codex or OpenCode agents in parallel with tavya: one task per agent, each on its own cloud computer and branch, so they cannot collide.

Why they cannot collide

Each task runs in its own world: a cloud computer with its own copy of the project's repositories, working on its own branch. Agents share nothing while they work. Their changes meet only when they land, one at a time.

Start several

Several tasks. Press New task (or N), say what you want, and press Ctrl+Enter (⌘+Enter on a Mac). Do it again for the next one; each starts at once. From a project's folder, the tavya CLI does the same:

tavya task new "Fix the login bug"
tavya task new "Add rate limiting to the API"
tavya task list

Several attempts at one task. In the task form's side column, set Attempts (1 to 8). The same task then runs that many times, each attempt on its own computer. When you confirm one at Review, tavya asks Keep or cancel other attempts? On a task's page, + New attempt adds another later, as a draft you can edit first.

Sub-tasks. An agent can split its own task into sub-tasks, each with its own agent. Ask for it in the prompt, for example "Split this into sub-tasks, one per package." A sub-task uses its parent's model and computer size unless the agent chooses others. Its work lands on the parent's branch, the parent's agent reviews it, and you review the parent once. The task's page lists them under Sub-tasks.

What limits how many run at once

  • Your plan. "Agents at once" is how many can work at the same time; more wait in line. See Pricing.
  • Model logins. A Claude or ChatGPT subscription has usage windows. When every login is busy or out of quota, a task says so, for example Waiting for a free Claude login or Waiting for Claude quota. Add more accounts under Settings → Agents; tasks use the first available one.
  • Computers. With the provider on Any, a full account gives way to the next, and a task waits only when all are full. Your own cloud account has its provider's limits (Cloud computers).

The project's Queues tab shows Agent work: the agents running now, then the waiting work in the order it will start, higher priority first. Drag a row to change the order. A task that says Held for higher-priority work has a Run now button.

How results come back

Each task stops at Review on its own. The For me view of the task list shows the tasks that wait for you. Read the changes, then Confirm, or reply with what to change.

Confirmed tasks move to Landing. Their changes merge one at a time through the Merge queue on the Queues tab, in an order you can drag.

Mix models

The model is chosen per task, in the model field under Agent in the task form. One task can run Claude Code, the next Codex, a third an open model through OpenCode. Settings → Task defaults sets what new tasks start with.

Next