Settings, cost and limits

Steps limit

A step is one decision plus the tools called for it. An agent that reads mail, thinks, and sends one reply uses about three. The limit exists to stop a confused agent looping forever at your expense; hitting it ends the run as a failure.

Ten suits most agents. Raise it when a run legitimately handles many items - triaging twenty emails needs more steps than triaging two. If an agent keeps hitting the ceiling, that is usually a sign the job is too broad rather than the limit being too low.

Daily spending limit

A cap on how much one agent may spend in a day. It is per agent - each has its own allowance, and a busy agent cannot exhaust a quiet one's budget. The count resets at midnight UTC.

The check happens before each step, so a run that starts under the cap can finish slightly over it. Treat it as a guardrail against runaway loops, not an exact accounting limit.

Think step by step

Asks the model to reason through each decision before acting. Slower and slightly more expensive, and worth it when the agent has to exercise judgement over ambiguous text. For simple routing it mostly adds cost.

Web search

Gives the agent a search tool. No account or API key is needed. The results setting controls how much comes back on each search:

Read full pages adds a second tool that opens a URL and reads it, for when a snippet is not enough. It is slower and much heavier on tokens, which is why it is separate. Long content is truncated.

It reads data files as well as pages. If a site publishes a CSV, JSON, or XML feed, point the agent at that instead of the page — it is cleaner, more complete, and far less likely to break when the site is redesigned. When the agent does read a page, links are kept alongside their text, so it can follow a list of results through to the individual entries rather than only seeing their titles.

Remember past runs

The agent finishes each run by writing itself a short note — which items it acted on, what it decided to skip — and the next few runs start with those notes in front of them. Because the agent chooses what goes in the note, it can carry forward exactly the identifiers it needs to avoid doing the same work twice.

Notes are kept from the last five successful runs, newest first, up to a fixed size — so an agent that writes a lot pushes out its older notes rather than everything getting squeezed. Useful for agents that should notice a pattern across days or must not repeat themselves; unnecessary for agents where each run stands alone. Costs a small amount on every run.

What things actually cost

• AI tokens go to OpenAI or Anthropic on your own key, at their prices. The platform does not mark anything up.
• SMS goes to Twilio, per message.
• Gmail, Telegram and web search cost nothing.

The dollar figures shown around the app are estimates from token counts at typical prices. Your provider's bill is the real number.

Notifications

Setod can alert you when something needs your attention. Go to Workspace → Integrations → Notifications to configure this.

• Email is always on. Alerts go to your login email address as a fallback for every event.
• Telegram — connect a Telegram Account connector and select it in the Notifications card. Alerts arrive in your Saved Messages, so they only you see them.

Events that trigger an alert:

• An agent is waiting for your approval before it can act.
• A run fails — sent at most once per hour per agent to avoid noise.
• An agent hits its daily spending limit.
• A connector is revoked and its tools have stopped working.

The Send test button on the Notifications card fires a test through every configured channel so you can confirm delivery before anything real happens.