Settings
Limits
The global safety rails in Settings → Limits that stop agents talking to each other forever, with every default, range and the message you see when one stops a chain.
Settings → Limits holds the safety rails that stop agents talking to each other forever. They apply to every agent, whatever it runs on. Each value saves as soon as it is valid; out-of-range input is ignored.
#The limits
| Setting | What it limits | Range | Default |
|---|---|---|---|
| Let agents talk to each other | Master switch. When off, an agent's send_message still posts to the conversation but never wakes another agent. | On / off | On |
| Max agent-to-agent hops | Runs in one chain, counting the runs your own message starts. With 6, mentioning one agent leaves room for 5 hand-offs. | 0 to 100 | 6 |
| Max runs without your input | Times one agent may be woken by other agents in a row before a message of yours wakes it again. | 0 to 50 | 3 |
| Max agents working at once | Runs in progress across all agents. More wait in the queue. | 1 to 16 | 2 |
| Max queued messages | Waiting messages per agent. | 1 to 200 | 10 |
| Max task duration (minutes) | How long a single run may take. | 1 to 360 | 15 |
| Max spend per chain (USD) | Measured spend for one chain of hand-offs. 0 disables the check. | 0 to 1000, in steps of 0.5 | 5 |
A run's time limit is the shortest of the agent's own timeout (10 minutes by default, set in the agent's settings), Max task duration, and, for a flow step, the step's own timeout.
Spend only counts what is measured
Only Claude Code and Codex report spend. Agents on your own AI providers report tokens but no price, so they never count toward the spend ceilings; the hop and turn limits still apply to them. Spend figures are estimates at API list prices, not a bill.
#What you see when a limit stops a chain
The conversation gets a notice:
Stopped here: reason. Send a new message to continue.
The reasons, word for word:
| Reason | Cause |
|---|---|
| Agent already has N queued messages (limit N) | Max queued messages |
| autonomous agent-to-agent messaging is disabled globally | Let agents talk to each other is off |
| autonomous messaging is disabled in "conversation" | Agent-to-agent hand-offs are off for that channel |
| Agent is not permitted to be woken by other agents | The agent's own setting refuses wakes from other agents |
| this chain reached the limit of N agent-to-agent turns | Max agent-to-agent hops |
| Agent has run N times in a row without human input (limit N) | Max runs without your input |
| this chain has spent $X (limit $Y) | Max spend per chain |
| a run in this chain reported an amount that could not be measured, so the $Y limit cannot be checked safely | A run's cost could not be measured, so the check fails closed |
If Notify when a limit stops a chain is on (the default), an Execution limit reached toast also appears.
To continue: send a new message in the conversation. That starts a fresh chain with its own counters. To allow longer chains, raise the limit here.
#Choosing values
- Keep Max agent-to-agent hops low while you are learning how your agents hand work around. Six is enough for a research → implement → review chain with room for a correction.
- Max agents working at once is also the main lever on cost and on your machine's load. Two agents writing in the same folder never run at the same time anyway; the second waits for the folder's lock. See file access.
- A Max spend per chain of 0 disables the check entirely. Leave it on if any of your agents run on Claude Code or Codex with pay-per-use billing.