Background agents and a supervisor: work that runs on its own stays within limits

Background agents wait in line; agents you open yourself never count against these limits or wait.

• At most 3 background agents at once (adjustable from 1 to 12).
• At most 40 new background agents a day (1 to 500).
• Usage threshold of 85%: when the 5-hour usage of an account background agents use passes it, jobs wait. 50 to 95, or no limit.
• Priority: merge queue first, then automations, then parallel teams.
• Background agents start on Automatic model by default: each job is rated simple, normal or hard and goes to that model.

Catching stuck agents is on by default and watches for three things: the same shell command 5 times within 15 minutes, the same error 3 turns in a row, and 20 minutes without a file change. The thresholds can be changed.

• Step one: the agent is redirected; a looping turn is stopped first.
• Step two: an agent on Automatic model moves to the hard-tier model and keeps its conversation.
• Last step: you’re told through a desktop and Telegram notification.
• Only background agents are watched by default; choose all to include your own.

With “Ajanlar gerektiğinde yeni ajan açabilsin” on (the default), Claude Code and Codex agents open, wait for, message and stop helper agents through the local agentvera-agents MCP server’s spawn_agent, wait_agent, agent_status, message_agent and stop_agent tools.

• Helpers appear under their parent agent with an “oto” badge, and you can step into them.
• A helper that changes code can work in its own worktree on an oto/… branch.
• One agent can open at most 3, and at most 6 can run in total; helpers can’t open agents themselves.
• Stopping a helper once it hands in its work is on by default.
• Only for agents on this computer, not SSH or remote agents.