Day 71: Token Budgets per PR and per Engineer

If engineers never see the bill next to the diff, agents will maximize tokens by default. Publish per-PR and per-engineer token budgets in the PR UI and fail soft/hard when exceeded.

⚡ TL;DR: Attribute tokens to pr, user, run_id. Show cost estimate on the PR. Soft-warn at 80%, hard-stop at 100% of budget.

Attribution

# metering.py
def record(usage: dict):
    emit_metric("AgentTokens", usage["total_tokens"], dims={
        "pr": usage["pr"],
        "user": usage["user"],
        "model": usage["model"],
    })

// pr comment
export function budgetComment(used: number, cap: number, usd: number) {
  const pct = Math.round((used / cap) * 100);
  return `Agent budget: ${used}/${cap} tokens (~$${usd.toFixed(2)}) — ${pct}%`;
}

Soft vs hard

def enforce(used: int, cap: int):
    if used >= cap:
        raise RuntimeError("hard_budget")
    if used >= int(cap * 0.8):
        warn("soft_budget")

❌ Unlimited agent retries on a WIP PR with no owner-visible meter.

Closing checklist

• [ ] Attribute tokens to PR + engineer
• [ ] Comment budgets on PRs
• [ ] Soft warn + hard stop
• [ ] Monthly engineer rollups
• [ ] Separate budgets for eval CI vs interactive

Series navigation

Day 70: Project: Secure Coding-Agent Sandbox on ECS Fargate · Day 72: Prompt Caching Pitfalls

Last updated September 11, 2026