Skip to main content
Not yet released. This guide describes implemented changes awaiting rollout. Availability requires the corresponding backend and dashboard release.

​
Goal

Cap what an agent can spend, stop it instantly, and know what stops it on its own.

​
Prerequisites

  • An agent. Budgets come from budget: in stackshift.yaml and can be changed in the agent’s settings or with stackshift agents budget set.

​
Workflow

1
Set the AI budget for a month, the AI budget for one run, and the running-time budget for a day.
2
Watch the Spend tab: AI credits and running time per day against each budget, in Naira or US dollars.
3
When you need to halt the agent, press Stop on any of its pages, or run stackshift agents stop <agent>.
4
To go again, press Start. After a budget stop or an abuse stop, only the owner can start it.

​
The three budgets

Raising a budget is for the owner. Lowering one is a safe change. Your account’s AI credits and wallet still apply on top: the agent can never go negative.
  • AI per month: you are told at 80%. At 100% model calls are refused and the agent stops itself.
  • AI per run: at 100% that run’s model calls are refused and the run ends.
  • Running time per day: at 100% the agent sleeps with waking turned off until tomorrow or until you raise it.

​
What Stop does

You can stop any agent at any time. So can Stackie and connected agents, because stopping is a safe change with a one-step undo. Starting again after a budget or abuse stop is for the owner.
  • Refuses every model call and every action from the agent.
  • Turns off its inbox sending and its agreement limits.
  • Puts it to sleep with waking turned off.
  • Deletes nothing. Start undoes it.

​
What stops an agent on its own

  • Every call: max_tokens is required and capped.
  • Every run: the run budget, a time limit of 30 minutes by default, and at most 200 model calls.
  • A loop: the same request, the same prompt and tools, more than 5 times in 2 minutes in one run pauses the agent.
  • Five failures in a row: five failed or refused actions, or failed runs, pause the agent.
  • A spend spike, a flood of blocked internet requests, and many refused actions in a row raise alerts that can stop the agent.
  • The month: at 100% of the monthly AI budget the agent stops.

​
Expected result

The agent can never spend more than its budgets or your account holds, and Stop halts its model calls, actions, email sending and agreements within seconds.

​
Common failures

  • The agent says it stopped for its monthly AI budget. Raise the budget, then start it. Raising a budget is for the owner, signed in.
  • The agent sleeps and does not wake until tomorrow: it used its running-time budget for the day.
  • The agent stopped after failures: five failed actions or runs in a row. Read the last runs for the cause before you start it.

Model access for agents

Your agent calls Anthropic or OpenAI through StackShift with a short-lived token. No provider key is ever in the agent.

Your agent’s own actions: risk, approvals and levels

Declare the tools your agent has, give each a risk, and let StackShift apply the agent’s level, approvals and untrusted-text checks before they run.