A single chat completion costs a fraction of a cent. An AI agent that plans, calls three tools, reasons about the results, and produces a final answer isn't 4x that cost — it can be 10-100x, and the exact number is genuinely hard to predict in advance. That unpredictability, not the per-call price, is what actually blows up an agentic AI budget.
1. Agent cost compounds through context accumulation (each step resends prior history), tool overhead (schema tokens on every call), and retries (uncertain steps trigger re-attempts).
2. None of these show up in a simple "cost per request" model — they only show up once you trace an actual multi-step run.
3. A hard step limit and cost-per-run ceiling, enforced at the orchestration layer, is the single most effective guardrail.
Why a 5-Step Agent Isn't 5x a Single Request
The intuitive mental model — "an agent makes several LLM calls, so it costs several times what one call costs" — undercounts by a wide margin, because it treats each step as an independent, similarly-sized request. In practice, three compounding factors make agent cost grow faster than linearly with step count.
Factor 1: Context Accumulation
Every step pays for every prior step
Most agent architectures pass the full conversation history — the original task, every reasoning step, every tool call and its result — as input to the next step. This means input token count grows with each step, not just output. By step 5 or 6, the model is re-reading everything that came before, every single time, even though the actually new information at that step might be small.
Illustrative shape of how input tokens grow across a single agent run:
The output per step might stay roughly constant, but input cost compounds — a 5-step run's total input tokens across all steps is considerably more than 5x a single step's input, because later steps are paying for the full accumulated history, not just their own contribution.
Factor 2: Tool Call Overhead
Every available tool is paid for, whether or not it's used
Giving an agent access to tools means the model needs to see each tool's name, description, and parameter schema on every call where tool use is possible — this is fixed overhead, independent of whether that specific tool ends up being called. An agent configured with 10 tools pays that schema overhead on every single reasoning step, even in a run that only ends up using 2 of them.
On top of the definition overhead, every tool call that is made adds its result back into the context for the next step — a tool that returns a large payload (a full document, a large API response) compounds the context-accumulation problem from Factor 1 even faster.
Factor 3: Retries and Uncertain Steps
The multiplier that's hardest to predict
An agent that isn't confident in a tool result, hits an ambiguous instruction, or gets an unexpected response from a tool will often retry that step — sometimes explicitly, sometimes by looping through additional reasoning to work around the uncertainty. This is the single largest source of the 10-100x range: a well-behaved run might complete in 4-5 steps, while a run that hits an edge case can loop far longer before either succeeding or hitting a limit.
The failure mode that matters most: an agent with no hard step limit, stuck in a loop where a tool keeps returning a result the agent doesn't know how to act on, retrying indefinitely. Without a ceiling, this doesn't fail loudly — it just keeps running and keeps costing money until someone notices the bill, not the behavior.
Want to estimate cost per agent run before you ship?
Model costs + tool calls + retries, in one spreadsheet — plug in your agent's shape and see the range.
Get the Agent Cost Estimation WorksheetEstimate cost per agent run. Model costs + tool calls + retries. One spreadsheet.
The Guardrails That Actually Bound This
Hard step limit. A maximum number of steps per run, enforced by the orchestration code, not left to the agent's own judgment about when it's done. This is the single most effective cap on the retry-driven tail of the cost distribution — it turns "could run indefinitely" into "will stop here regardless."
Cost-per-run ceiling. Independent of step count, a hard dollar or token ceiling per run catches the case where a small number of steps each involve unusually large context or tool payloads — a scenario a step-count limit alone wouldn't catch.
Context trimming between steps. Rather than passing the full unmodified history to every step, summarizing or truncating older steps once they're no longer directly relevant reduces the input-token growth from Factor 1 — the tradeoff is some risk of losing context the agent actually needed, so this needs testing against your own task, not a blanket policy.
Budget enforcement at the proxy layer. Since an agent's calls still route through the same LLM proxy as everything else, the same hard-block budget pattern that works for runaway single-request loops applies here too — a workspace-level ceiling that doesn't depend on the agent's own logic behaving correctly.
Why This Deserves Its Own Cost Model
Budgeting for agentic AI using a per-request cost estimate — "each call costs about $X, we'll make roughly Y calls" — systematically undercounts because it misses all three compounding factors above. The realistic way to budget is to actually trace a handful of real agent runs across the full range of tasks you expect (simple, typical, and edge-case), measure the actual step count and token accumulation for each, and use that range rather than a single-call estimate multiplied by an assumed step count.
This is also why per-user cost visibility matters even more for agentic features than for simpler ones — a small number of users hitting the edge-case-retry tail of the distribution can dominate total cost in a way that's invisible in an aggregate spend number.
Frequently Asked Questions
Why do AI agents cost more than a single LLM request?
What's the biggest hidden cost in agentic workflows?
How do tool calls affect agent cost?
What's the single most effective control for agent cost?
Estimate your agent cost before it surprises you.
Model costs, tool call overhead, and retry multipliers — in one worksheet, so you can budget for the range instead of a single-call estimate.
Get the Agent Cost Estimation WorksheetFree forever up to 10K requests. No credit card required.
