An agent mid-task doesn't get to pause and go find a credit card. If it's three calls into a ten-call job - drafting a reply, sending it, polling for the next message - and its wallet balance runs out on call four, the whole run stalls on an authentication failure that has nothing to do with the task itself. Auto-reload exists specifically to keep that from happening.
Auto-reload is a threshold and an amount, both of which you set on the wallet at /wallet: once the balance drops below the threshold, DimeBox charges the saved card for the amount and credits the wallet, without anyone having to notice the balance was low first. The agent calling the API never sees this happen - from its side, the wallet just keeps having enough balance to authenticate the next call.
A human using a metered API notices a low-balance email and tops up manually. An unattended agent - a cron job polling an inbox, a loop that drafts and sends follow-ups on its own schedule - has no equivalent moment of noticing anything. Building a Cron-Triggered Agent That Checks Its Own Inbox and Adding Email to an Autonomous Agent Loop both assume the loop runs without a person watching it; auto-reload is what makes that assumption safe to make on the payment side too.
The two numbers don't have to match your spend pattern precisely, but they should bracket it: a threshold too close to zero means a slow top-up charge might not clear before the next call needs the balance; a threshold too high means you're reloading more often than necessary. A reasonable starting point is sizing the threshold to cover a few hours of expected call volume at whatever mix of $0.02 drafts, $0.01 sends, and $0.03 priority receives your agent actually makes, and sizing the reload amount to cover a few days of that same volume so you're not reloading on every other call.
Auto-reload refills the balance; it doesn't raise what the wallet is allowed to spend in a day. If a daily spend limit is set lower than what auto-reload would otherwise permit, the limit still caps what gets spent - see Setting Daily Spend Limits on an Autonomous Agent's Wallet for how the two settings interact. A reloaded wallet that's already hit its daily cap still can't spend more until the next day, regardless of balance.
A new wallet is capped at a lower balance and a lower daily spend limit for its first few days, regardless of what auto-reload is configured to do - see New-Wallet Risk Controls. If you're testing auto-reload on a brand-new wallet, expect the cap to bind before your configured amount does, until the wallet ages past that initial window.
Auto-reload uses the same saved card as the original funding. If that card is expired or declined when a reload attempt fires, the wallet doesn't get credited and continues operating on whatever balance it already had - which, if the reload was triggered by a low-balance threshold, might not be much. That failure mode, and how to design an agent that notices it instead of retrying blindly, is covered in What Happens When an Autonomous Agent's Card Payment Is Declined.
Auto-reload solves "don't run dry mid-task"; it doesn't solve "don't overspend." Those are separate problems with separate controls - a spend limit caps the ceiling, auto-reload keeps the floor from hitting zero unexpectedly. Budgeting for a Fleet of Autonomous Agents That Pay Their Own Way covers the broader planning question of how much balance a given population of agents actually needs, which auto-reload amount and threshold should follow from, not the other way around.
Auto-reload keeps the balance from hitting zero unexpectedly; it doesn't address the other two ways a wallet can stop authenticating calls. A daily spend limit caps total spend for the day regardless of how full the balance is - see Setting Daily Spend Limits on an Autonomous Agent's Wallet - so a wallet that's hit its limit will still reject calls even with a freshly reloaded balance sitting behind it. And a wallet DimeBox has frozen for suspected abuse won't start working again just because auto-reload successfully charged a card; that's a different kind of stop entirely. Treat auto-reload as solving exactly one failure mode - running out of funded balance mid-task - and keep separate monitoring for the other two.
Say an agent sends 50 follow-ups and drafts 50 replies a day: that's 50 times $0.01 plus 50 times $0.02, or $1.50/day in routine spend, before any priority receive calls. A threshold around a day or two of that ($1.50-$3.00) and a reload amount covering a week ($10-$15) keeps the wallet comfortably ahead of its own spend pattern without triggering a top-up on every other call.