# Setting a daily send cap for an email agent: how to pick the number

> Pick an email agent's daily send cap from your own data: run in Suggest first, count replies you approve unchanged, then set the cap just above a normal day.

- Published: 2026-11-07
- Author: Koltrix Team
- Category: AI & email
- Source: https://koltrix.com/blog/setting-a-daily-send-cap-for-an-email-agent-how-to-pick-the-number

A daily send cap is a hard limit on how many messages an email agent may send on its own in one day. Pick it by measuring a normal day while a person still approves every reply, then set the cap a little above that number, so a bad day stops early.

The cap is not a forecast of how busy you will be. It is a ceiling on damage. If an agent misreads a thread, loops with another bot, or answers a flood of near-identical messages, the cap decides whether that costs you five wrong emails or five hundred. This post walks through how to choose the number, what happens when the agent reaches it, and what the cap does not protect you from.

## What does a daily send cap do?

A daily send cap counts the replies an agent sends by itself and holds everything after the limit for a person to approve. In Koltrix, Auto mode is Beta, it is off by default for each agent, and it needs a daily cap greater than zero before you can switch it on. A cap of 0 means Auto sends nothing.

The count covers autonomous sends only. Replies a person approves in the Action center do not use up the cap, and neither do drafts saved in Suggest mode. The day is counted from midnight in the agent mailbox's own timezone, so set that timezone to where your customers or your team actually are, not to wherever the server sits.

## What happens when the agent reaches the cap?

When the cap is reached, the agent does not drop the message and does not keep sending. The reply is held and shows up in the Action center as a normal item waiting for a person, with the reason `daily_cap`. A person can still approve it, edit it or ignore it. Pending approvals expire after 7 days, and a mailbox can hold at most 100 open items, so an agent that hits its cap every day is telling you something (see the signals section below).

This matters for choosing the number. A cap that is too low costs you speed, because some replies wait for a tap. A cap that is too high costs you safety. Those costs are not symmetric: a held reply can be sent a minute later, while a wrong email that already left cannot be recalled. When unsure, start low.

## How do you measure a normal day first?

Run the agent in Suggest, or in Approve, for at least a week before you turn on Auto. Suggest only saves drafts. Approve proposes a send and waits for a person. Neither sends anything alone, and both give you data. Our guide to [what an approval step should show](/blog/what-an-approval-step-for-ai-email-should-show-and-when-it-expires) covers how to review them well.

For each day, write down three numbers:

1. How many messages the agent could have answered.
2. How many of its replies you would have sent unchanged.
3. How many you edited heavily or rejected.

Only the second number is a candidate for Auto, because Auto only helps with replies you would have approved anyway. If the agent drafts forty replies a day but you send five unchanged, the useful cap is nearer five than forty.

## How do you turn those numbers into a cap?

Take the largest count of "sent unchanged" you saw on an ordinary day in the measuring week, ignore the one freak day if you can name its cause, and add a small margin. A table makes the logic easy to apply. The numbers below are examples to show the arithmetic, not recommendations for your volume.

| What you observed in a week | Reading | Example starting cap |
|---|---|---|
| Unchanged replies per day: 2, 3, 3, 4, 2 | Steady and small | 5 |
| Unchanged replies per day: 8, 11, 9, 10, 31 | One spike, caused by a known incident | 12 |
| Unchanged replies per day: 0, 1, 0, 2, 1 | Too little traffic to learn from | 3, or stay on Approve |
| Most replies edited before sending | Agent is not ready | Stay on Suggest |

When you first switch an agent to Auto in the Koltrix app, the cap field is filled with 20 if it was empty. Treat that as a placeholder to overwrite, not a recommendation. Twenty may be far too many for an agent that answers five invoice questions a day.

Two habits keep the cap honest:

- **Round down, not up.** If the data says 9, set 10, not 25.
- **Review it on a schedule.** Check the cap after the first week on Auto, then monthly, or whenever your volume changes.

## Which signals say the cap is wrong?

Look at how often the agent hits it and what is waiting.

- **Hit every day before noon.** Either volume grew, or something is generating replies you did not expect. Read the held items before raising the number.
- **Never above a third of the cap.** The cap is loose, so it protects little. Lower it toward the real peak.
- **Many held items with reason `daily_cap` that you approve unchanged.** The agent is doing good work and the cap is too tight. Raise it in small steps.
- **Many held items you reject.** Do not raise the cap. Go back to Suggest and fix the agent's notes or scope.

## What does the cap not protect against?

A cap limits quantity. It says nothing about quality, so it works alongside other rules and does not replace them. In Koltrix, Auto also requires every one of these:

- High confidence (at least 0.85) and an answer grounded in the mailbox's own notes, documents or facts.
- Recipients who are already in the conversation, and an optional allowlist that can narrow it further.
- A message being answered that the injection shield rates as no risk.
- Hours outside any quiet hours you set, a 10-minute hold on identical messages, and a kill switch that puts every agent back to Suggest.

A cap of ten still lets ten wrong answers through if all ten pass those checks, so keep the agent's scope narrow. Our posts on [least privilege for email agents](/blog/least-privilege-ai-email-agents) and [safe tasks for email agents](/blog/ai-email-agents-safe-tasks) explain how to choose what an agent should handle at all. If two bots can answer each other, a low cap is also your last line of defense against a loop, which we cover in [when two agents answer each other](/blog/when-two-agents-answer-each-other-stop-email-loops-between-bots).

## A short checklist before you turn Auto on

- The agent has run in Suggest or Approve for at least a week.
- You know your count of replies sent unchanged on a normal day.
- The cap is that count plus a small margin, not a round number you liked.
- The mailbox timezone is correct, because it sets when the day resets.
- You know where held items appear, and who checks the Action center.
- You know where the kill switch is and who is allowed to use it.

For the background on giving an agent its own address, see [give an AI agent its own email address](/blog/give-an-ai-agent-its-own-email-address-what-it-takes-and-what-can-go-wrong), and the [agent mailboxes feature page](/features/agent-mailboxes) for the controls themselves.

## Bottom line

Choose the daily cap from evidence: a week of Suggest or Approve, a count of replies you would send unchanged, and a small margin above a normal day. Set it low, watch how often the agent reaches it, and raise it in small steps only when held replies are ones you keep approving. Auto is Beta and Suggest remains the default, so start there, and treat the cap as the limit on how wrong a bad day can get.