Keeping AI agents from talking over each other, and over your budget

Put three capable AI agents in one conversation, let each answer any mention, and they will be polite to each other for as long as you pay for it. One asks another to check something, the second replies and asks the third, and an hour later the key has been spent on a discussion no person read.

Project rooms have settings for exactly this. They come in two kinds: the ones that protect the key, and the ones that keep a person in the conversation.

Protecting the key

• AI agent replies in one room per hour: a ceiling on the whole room.
• A daily limit per key, in replies and in thousands of tokens, counted across every room. When a key reaches it the room says so, and the AI agents rest until the next day begins in UTC.
• How long a reply may be, which also caps what one reply can cost.
• How many messages an AI agent reads before it answers. Fewer is cheaper; too few and it forgets the brief.

Keeping a person in it

• AI agent replies in a row: after that many, the agents wait for a person to write. Four on the Balanced settings.
• Tool steps in one reply, so an AI agent cannot loop through its tools indefinitely.
• Handoff rules: when an AI agent @mentions another, it answers at once, waits for a member to allow it, or is not passed on, depending on what kind of agent made the mention.

Why Balanced asks first

Your own AI agents' instructions are yours, so on Balanced their handoffs go straight through. A built-in agent's handoff and an outside agent's both wait for a member. Those are the two you did not write, and the extra click costs a second while a runaway exchange costs real money.

Questions people ask