How to Set Per-Agent Budgets for Your AI Fleet

You have ten agents running in production. Maybe fifty. Each one calls LLMs on your behalf, burns tokens, and adds to a bill that arrives once a month with no line-item breakdown by agent.

Per-agent budgeting fixes that. It gives every agent in your fleet a spending allowance, forces anomalies to surface before they become invoices, and lets you make real decisions about which agents earn their keep.

This is not complicated in theory. In practice, most teams skip it because nothing in the standard LLM API toolchain makes it easy. Here is the full approach.

Why Per-Agent Budgets Matter

Your fleet is not homogeneous. A summarization agent running on haiku to process 200-word emails is nothing like a research agent running on GPT-4o to analyze 50-page PDFs. Treating them as the same cost center is how you end up with $40,000 bills that nobody can explain.

Per-agent budgets matter for three reasons:

Accountability. When every agent has a named budget, cost suddenly has an owner. The team that ships a new agent is also the team responsible for that agent's spending. That changes the incentive structure—engineers start asking "how much does this agent cost per task?" before they ship.

Early warning. A flat $10,000/month LLM budget tells you nothing useful until the last week of the month. A per-agent budget of $800/month on your email-triage agent sends an alert on day 12 when it hits $640. That is actionable. You can investigate before the problem compounds.

Model selection. Once you know what each agent actually costs, you can make rational decisions about which ones justify expensive models. Your executive-summary agent might earn GPT-4o. Your ticket-classifier absolutely does not.

How to Calculate Agent Cost Baselines

Before you can set a budget, you need a baseline. The baseline answers: *what does this agent normally cost per day, per task, and per user?*

Start by instrumenting three numbers for each agent:

Average tokens per task. Input tokens plus output tokens, averaged over at least 100 real tasks. This is your cost denominator. If you do not have 100 tasks yet, use 20 and mark the baseline as provisional.

Tasks per day. From your logs or scheduler. Multiply by the daily active user count for user-triggered agents.

Model price. Check the current pricing page for every model your agent uses. If the agent can switch models, log which model each call used.

The formula:

```

daily_cost = (avg_input_tokens × input_price_per_token)

+ (avg_output_tokens × output_price_per_token)

× avg_tasks_per_day

```

Do this for each agent. Plot it in a spreadsheet. You will immediately see outliers—agents that cost 10x more per task than you expected.

Common surprises when teams do this exercise:

• Agents with retry logic cost 3–8x their nominal per-call price because failures are invisible in billing
• Agents that accumulate conversation history grow linearly in cost per task as history length increases
• Scheduled agents running during off-hours account for 30–60% of total spend at companies with large fleets

Once you have baselines, set your budget at 1.5× the 90th percentile daily cost. That gives you headroom for legitimate spikes while still catching runaway behavior.

Setting Alerts vs. Hard Limits

These are different tools for different problems. Use both.

Soft Alerts (80% threshold)

Trigger at 80% of the monthly budget. The alert should:

• Fire to Slack or PagerDuty (not just email)
• Include the agent name, current spend, remaining budget, days left in period
• Be non-blocking—the agent keeps running

This is your "something might be wrong" signal. Most months, you will investigate, find nothing unusual, and move on. But when you find a retry storm or a context-accumulation bug, you will find it with 20% budget headroom rather than a $0 balance.

Hard Limits (100% threshold)

When an agent hits its budget ceiling, you have three options. Pick based on the agent's role:

Stop. Best for background/scheduled agents that are not user-facing. Queue tasks for the next billing period. Simple, safe, prevents further spend.

Degrade. Switch to a cheaper model for the remainder of the period. A claude-3-5-sonnet agent degrades to claude-3-haiku at the limit. Users see slower/lower-quality output but service continues.

Alert-and-continue. For mission-critical user-facing agents where stopping or degrading breaks the product. The agent keeps running, but you get a P1 alert and the cost goes to an "overage" bucket tracked separately. Use sparingly.

The wrong move is to silently continue with no limit enforcement. That is the default behavior of every LLM API today—and it is how teams end up with five-figure surprises.

The SpendPilot Approach

Per-agent budgeting is the core of what SpendPilot is built for. Here is how it works in practice:

Agent tagging at the API layer. Every LLM call your agents make gets tagged with an agent identifier before it hits the provider API. No changes to your agent code required—tagging happens in the proxy layer.

Real-time cost dashboard per agent. Not an end-of-month report. A live dashboard showing current spend, budget remaining, and burn rate for every agent in your fleet. If something is accelerating, you see it now.

Configurable budget policies. Set monthly or daily caps per agent. Choose the enforcement action (stop, degrade, alert) per agent. Policies are stored in config, version-controlled, and apply immediately without redeploy.

Automatic anomaly detection. SpendPilot compares current burn rate against the rolling 14-day baseline for each agent. If an agent starts spending 3× its normal rate, you get an alert before it hits 50% of its budget.

Token-level attribution. Every call logged with input tokens, output tokens, model version, agent ID, task ID, and user ID. When you get a spike alert, you can trace it to the exact task and user within 60 seconds.

Engineering leads managing 10–100 agents typically see two outcomes within the first month: they discover one or two agents costing far more than expected, and they reclaim 20–40% of their LLM budget by right-sizing those agents.

The math is straightforward. If you are spending $8,000/month on LLMs and 30% is going to agents that could run on a cheaper model, that is $2,400/month in recoverable spend. SpendPilot pays for itself in the first week.

See your agent spend in real time → spendpilot-3.polsia.app