The Loop Tax: Why AI Agents Are So Expensive
These figures are from our June 2026 snapshot (210,840 calls). The refreshed July corpus (285,814 calls) shows the same pattern more sharply: 97% of spend in the orchestration loop. See the updated teardown.
Agents are expensive mostly because of the loop, not because frontier models cost more. We metered 210,840 governed tool calls across our own workspaces as of June 2026. The orchestration loop was 89% of the spend. Here is how it works and what moves it.
Leaf calls and loop calls
A leaf call is what most people picture: one prompt in, one answer out. Summarize this. Classify that. The cost is predictable.
An agent is a loop, the engine of its harness. The model picks a tool, reads the result, picks the next tool, reads that result, and repeats until it is done. On every turn it reads the whole growing context again: the system prompt, the tool definitions, every earlier tool result, and the running transcript.
So a 12-turn agent is not 12 cheap calls. It is 12 calls where the input grows each turn, and you pay for all of it every turn. That repeated reading is the loop tax. Cost scales with the number of turns, not the number of useful answers.
We tag every governed model call as callKind: loop (the model orchestrating tools, turn by turn) or leaf (a single sub-task). Split a real workload that way and the loop dominates. It was 89% of spend, and the orchestration step (chat.completion) averaged 10.3 seconds. The loop is the slow part and the expensive part. The one useful answer at the end costs little next to the trip to get there.
Why a rate limit won’t save you
When the bill spikes, the first instinct is a rate limit. It doesn’t help. A rate limiter caps how many calls you make, not what each call costs. A 12-turn loop on a frontier model is well inside a call-count budget and still expensive, because each turn pays the re-read. Call-count limits can’t see how big each call got or which model answered it. We covered that failure in stop your agent from burning your API budget.
Three levers
1. Put a cheaper model on the loop
This was the biggest lever in our data. In the teardown, one frontier model was 80% of the bill across 7.6% of the calls. That is roughly 114 times the per-call cost of the cheap model that did most of the work. Most orchestration turns (“which tool next?”) don’t need a frontier model. Run the loop on a fast, cheap model such as Flash or Haiku, and keep the expensive one for the leaf calls that need real reasoning. We did not measure the quality change in this dataset, so test it on your own tasks before you switch.
2. Cache the stable prefix
The system prompt and tool definitions are the same on turn 12 as on turn 1, yet you pay for them every turn. Prompt caching bills that repeated prefix at the cache price. Cache reads cost roughly a tenth of full price. The catch is knowing whether caching is working. ACP meters the cached portion of each call it prices, so you can read the number instead of hoping a setting is on.
3. Cap the loop
A runaway loop is what turns a $0.04 run into a $4 one. The agent gets stuck retrying or wanders off task. The fix is a ceiling on the loop itself, such as a turn limit or a per-agent budget, set before the run instead of a monthly cap you find out you blew afterward. Check what your harness supports. ACP gives you the numbers to pick a sensible ceiling: turns per run, cost per run, and which agent ran away.
You need the split first
All three levers assume you can see the split: loop or leaf, which model, which agent, how many turns. A monthly invoice from your model provider can’t tell you whether you are paying for orchestration or for work, or which agent’s loop is the problem. ACP records the model, the callKind, the estimated cost and the agent on each call it meters, so you can measure the loop tax and then cut it. This is AI agent cost monitoring.
To see your own loop-and-leaf split, sign in. The cost X-ray is in the dashboard. For how the metering works, start with what an agentic control plane is.
Cost figures are a 2026-06 snapshot from our own internal workspaces, not customer data.
Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:
curl -fsSL https://agenticcontrolplane.com/install.sh | bashirm https://agenticcontrolplane.com/install.ps1 | iex