Stop every agent run at fixed limits set in code

How to do it

A loop of model and tool calls can run for hours when a tool fails or a page tells it to retry. Keep the limits in settings, per agent, and read them in the loop before each step. Set a maximum output length on every model call, and rate limit the chat endpoint like any other. Count the cost of model calls and paid tools against the run and the user. Stop the run with its status when a limit is hit, as end every agent run with a known status says. The per-feature budget is in price every model call and cap the spend. Spend caps for outside agents that call your product are in Cap what each agent grant may spend. Caps on coding agents are in Cap what your coding agents can spend.

Examples

TypeScript example

const LIMITS = { maxSteps: 10, maxRunMs: 120_000, maxRunUsd: 1, maxUserUsdPerDay: 20 } as const;

export function limitReached(run: AgentRun, spentTodayUsd: number): RunStatus | null {
  if (run.steps >= LIMITS.maxSteps) return 'step_limit';
  if (Date.now() - run.startedAt.getTime() > LIMITS.maxRunMs) return 'time_limit';
  if (run.costUsd >= LIMITS.maxRunUsd || spentTodayUsd >= LIMITS.maxUserUsdPerDay) return 'budget_limit';
  return null;
}

Python example

LIMITS = AgentLimits(max_steps=10, max_run_seconds=120, max_run_usd=1.0, max_user_usd_per_day=20.0)

def limit_reached(run: AgentRun, spent_today_usd: float) -> RunStatus | None:
    if run.steps >= LIMITS.max_steps:
        return "step_limit"
    if time.monotonic() - run.started_at > LIMITS.max_run_seconds:
        return "time_limit"
    if run.cost_usd >= LIMITS.max_run_usd or spent_today_usd >= LIMITS.max_user_usd_per_day:
        return "budget_limit"
    return None

How to check

Tests run the agent against a fake model that never finishes and one whose tool calls each cost more than a tenth of the limit. The first stops at the step limit and the second at the spend limit. A search of the code finds no limit value set from model output.

Sources

• OWASP LLM06:2026 Unbounded Consumption

Related rules

• AID-24 Skip approvals only in a throwaway sandbox in AI-assisted development
• AIF-42 Price every model call and cap the spend in AI features
• AIF-52 Use a fixed workflow when the steps are known in AI features
• AIF-53 End every agent run with a known status in AI features
• AGT-25 Cap what each agent grant may spend in Agents as users
• AID-62 Cap what your coding agents can spend in AI-assisted development
• AIF-72 Give each sub-agent no more power than the agent that started it in AI features
• PRC-10 Let owners set a cap on each member and key in Pricing, plans and entitlements