The Agentic Loop Multiplier
Every "task" in a ReAct-style agent is not one API call. It is 4–10 — and each iteration bills against a growing context. Standard token calculators miss this by a factor of 4×–10×.
4×–10× — per-task token consumption in an agentic loop, vs. a naive single API call.
The definition
A single "task" in a ReAct-style agent framework is not one API call. It is a structured loop: plan, use tool, observe, reflect, act again. Each iteration is a billable API call with its own input context — which grows at every step as the agent appends its prior reasoning to the scratchpad. This is the Agentic Loop Multiplier.
Standard token calculators assume a 1:1 relationship between user tasks and API calls. In agentic deployments this assumption is false by a factor of 4× to 10×.
Why it costs money
Three compounding effects, all invisible in a naive tokens × tasks estimate:
- Iteration count. Each ReAct loop turn is a billed call. Complex tasks routinely take 4–10 turns. Multiply by hundreds of thousands of annual tasks and the API bill stops looking like the friendly demo.
- Context growth per turn. The input context to each turn includes the running scratchpad — prior reasoning, tool outputs, observations. Input tokens per call grow monotonically inside a single task.
- Abandonment tax. Tau-bench (2024) measured 12–22% task abandonment across state-of-the-art agents. Abandoned tasks still burn roughly half their token budget before failing. You pay for the failure.
The number
Applied consistently to a mid-level knowledge-worker deployment (250 daily tasks per user, GPT-4o class model, five-turn loop):
- Naive estimate: ~$28,000/yr in tokens
- Modelled with loop + abandonment: ~$110,000–$180,000/yr
- Multiplier: 4×–6× on the vendor slide
Longer-horizon agents (research, code, ops) push the multiplier toward the 10× end of the range. The full derivation, including the abandonment-cost formula and its Tau-bench sourcing, is in the methodology, Module 2B.
How to check it on your own deployment
- Instrument agent traces to record iterations-per-task and mean input tokens per iteration.
- Compute mean iterations × mean input tokens per iteration × input price + mean output per iteration × output price.
- Compare against the naive "prompt tokens × completions × price" line in the vendor deck.
- The gap is the loop multiplier. If it is under 3×, you are either running short-horizon single-call tasks or your traces are undercounting.
Why this framework is worth naming
Because giving the effect a name lets a CFO push back on a slide that shows only base token cost. "What is your agentic loop multiplier?" is a specific, answerable, budget-relevant question. "Is this token estimate realistic?" is not.
Model this on your deployment
This framework is baked into the DroidWork TCO calculator. Adjust the inputs — the numbers move accordingly.
Open the Calculator →Frequently asked
What is the agentic loop multiplier?
The factor by which per-task token consumption in a ReAct-style agent exceeds a naive single-call estimate. Typically 4× to 10×, driven by plan/act/observe/reflect iterations and by context growth as prior reasoning accumulates in the scratchpad.
Why is token cost 4× to 10× higher in agentic deployments?
Because a single user task triggers several API calls — one per loop iteration — and each successive call carries a longer input context that includes the prior reasoning and tool outputs. Abandoned tasks consume roughly half their token budget before failing, which adds an additional cost layer.
Where does the 4× to 10× range come from?
It is derived from combining measured iteration counts in production agents (typically 4–8 turns for complex tasks) with the empirical 12–22% abandonment rate measured in Tau-bench (2024) across state-of-the-art LLM agents.
How do I test this on my own deployment?
Instrument agent traces to capture iterations per task and mean input tokens per iteration, then compare the resulting cost against a naive "tasks × completion tokens × price" estimate. The gap is the multiplier for your workload.
See also
This framework is documented as part of the open-source DroidWork AI TCO Framework. Cite it as: DroidWork.ai (2026). "The Agentic Loop Multiplier." https://droidwork.ai/frameworks/the-agentic-loop-multiplier/