TL;DR: Self-running agent loops multiply token cost without multiplying yield. Anthropic's own research shows multi-agent systems burn roughly 15x the tokens of a chat, and token usage explains about 80% of performance variance. Context rot makes it worse: as the window fills, recall degrades, so longer loops produce weaker output. The fix is not more autonomy. It is bounded loops fed decision-grade context, so each pass starts with the full picture and earns its tokens.
The advice making the rounds from AI lab leaders in 2026 is seductive and wrong: build agents that detect their own data, write their own prompts, and keep the model running on a continuous loop. That guidance optimizes for the wrong variable. Continuous autonomy does not buy you efficiency. It mostly multiplies your AI agent token cost while output quietly degrades pass after pass. The headline question is simple, so here is the direct answer: self-running loops burn more tokens, not fewer, because agentic loops already cost multiples more than a single call, and context rot makes long loops less accurate as the window fills.
So the number that matters is not how long the model runs. It is token yield, the decision-grade output you actually get per token spent. Unbounded loops chase running time and ignore yield. That is how teams end up paying a rising AI agent token cost for results that flatten or decline. The discipline is the inverse of "keep it running." It is bounded loops fed curated, decision-grade context for agents, so each pass earns its tokens.
Why do self-running agent loops burn so many tokens?#
Self-running loops burn tokens because one request fans out into many model calls, not one. According to Anthropic's multi-agent research (Anthropic Engineering, 2025), agents use about 4x the tokens of a single chat, and multi-agent systems use roughly 15x. The same research found token usage alone explains about 80% of performance variance.
Here is the mechanism. A single chat is one prompt and one completion. An agent loop is different. It reads context, calls the model, invokes a tool, re-reads the result, calls the model again, and repeats until some stop condition fires. Each cycle re-sends accumulated history. So the AI agent token cost is not the sum of useful work. It is useful work times the number of passes, plus all the re-reading in between.
Anthropic's engineering team reports that agents consume roughly 4x the tokens of a single chat and multi-agent systems about 15x, while token usage explains around 80% of performance variance (Anthropic, 2025). That ratio reframes agentic token consumption as the dominant cost driver, not a rounding error.
Independent measurement makes the variance even starker. A study of how agents spend on coding tasks found that on SWE-bench Verified, agentic runs on the same task can differ by up to 30x in total tokens, and a full agentic run consumes on the order of 1000x the tokens of ordinary code chat. Worse, models badly misjudge their own spend: their self-predicted token cost correlates with actual cost at 0.39 or lower (arXiv, 2026). An unbounded loop is therefore betting on a cost the model itself cannot forecast.
A study of agent spending found that on SWE-bench Verified, agentic runs on the same task vary by up to 30x in total tokens, and full agentic runs consume roughly 1000x the tokens of ordinary code chat, while models self-predict their own cost at a correlation of just 0.39 or lower (arXiv, 2026). Unbounded loops have wildly unpredictable, often explosive cost.
What does "keep the model running" actually cost?#
"Keep the model running" costs more than its advocates admit, and the market is already pulling back. Per a Goldman Sachs analysis covered by Tom's Hardware (2026), agentic tools chain a single request into repeated calls, and Goldman projects token demand could rise on the order of 24x in the next few years if these patterns hold.
Notably, the same reporting describes Uber and Microsoft trimming per-token agent spend as tokenized billing bites. That is the tell. When the companies pushing agents hardest start metering them, the "always running" model is not the efficient default it was sold as.
The lab-leader pitch frames autonomy as savings. The bill says otherwise. Every extra pass re-pays setup overhead before it touches real work. Treat the 24x figure as a directional forecast, not a guarantee, but the direction is clear: agent loop cost compounds. The relevant question is whether each pass earns its keep, a question "keep it running" never asks.
And do not expect falling prices to bail you out. Epoch AI (Epoch AI, 2025) finds the inference price to reach a given capability level fell anywhere from 9x to 900x per year. Per-token price is collapsing, yet agent bills keep climbing. That tells you the runaway cost is driven by consumption, the loops, not by unit price.
Epoch AI reports that the LLM inference price to reach a fixed capability level dropped between 9x and 900x per year (Epoch AI, 2025). With unit price falling that fast, rising agent bills are a consumption problem. Unbounded loops, not expensive tokens, drive runaway cost.
Goldman Sachs, as reported by Tom's Hardware in 2026, projects agentic tooling could push token demand toward 24x current levels in the next few years, with Uber and Microsoft already curbing per-token agent spend (Tom's Hardware, 2026). Forecast, not fact, but the trajectory points up.
Why does running longer make agents worse, not better?#
Running longer makes agents worse because context is finite and recall degrades as the window fills. According to Anthropic's guidance on context engineering (Anthropic Engineering, 2025), context behaves as a limited resource with diminishing returns. As tokens accumulate, the model's effective attention budget thins, and recall of earlier instructions drops.
This is context rot, and it is the yield-killer. A long loop costs more per pass and produces lower-fidelity output per pass, because the window is noisier on iteration twenty than on iteration two. So you pay a rising AI agent token cost for results that are getting worse. That inversion breaks the "more autonomy equals more value" assumption at its root.
We have seen this play out in practice. Beyond a certain depth, agents start re-litigating settled decisions and looping on the same dead end. The "keep it running" crowd treats loop length as free upside. It is the opposite: each additional unbounded pass is a small bet against your own context window, and the house edge grows. Proof of the decay pattern lives in our context rot teardown.
More passes also do not buy more trust. The Stack Overflow 2025 Developer Survey (Stack Overflow, 2025) found 66% of developers cite "AI solutions that are almost right but not quite" as a top frustration, while only about 29% to 33% trust the accuracy of AI output. Looping the model again rarely closes that gap; it just spends more tokens circling the same near-miss.
The Stack Overflow 2025 Developer Survey found 66% of developers cite "AI solutions that are almost right but not quite" as a leading frustration, and only roughly a third trust AI output accuracy (Stack Overflow, 2025). Additional loop passes do not reliably convert near-right output into right output.
Anthropic's context-engineering guidance (2025) describes the model's usable context as a finite attention budget with diminishing returns, where recall degrades as the window fills (Anthropic, 2025). Longer self-running AI agents therefore trade rising cost for falling fidelity, the defining mechanic of context rot.
How do runaway loops compound context tax and context debt?#
Runaway loops compound two coined costs at once: context tax and context debt. In our own controlled test, documented in our same-prompt experiment, curated context cut tokens by 42% and tool calls by 64% against an identical model and prompt. That gap is what unbounded loops forfeit on every single pass.
Start with the context tax, the tokens spent loading tool definitions and preambles before real work begins, detailed in our MCP token budget autopsy. A loop re-pays that tax each iteration. Twenty passes means twenty rounds of overhead before useful tokens.
Then there is context debt, the accumulated cost of missing, stale, or contradictory context. When an agent writes its own prompts loop after loop, each self-authored instruction drifts a little further from the original intent. By iteration fifteen the agent is faithfully executing a goal nobody set. Two coined costs, one loop, both compounding silently while the token meter runs.
An Unblocked controlled test found that feeding curated context to an identical model and prompt cut token usage by 42% and tool calls by 64% (Unblocked, 2026). Runaway loops re-pay the context tax and accrue context debt on every pass, forfeiting that efficiency repeatedly.
Frequently asked questions#
How much more do AI agents cost than a single chat call?#
Substantially more, because one request becomes many model calls. Anthropic's engineering research reports agents use roughly 4x the tokens of a single chat, and multi-agent systems about 15x, while token usage explains around 80% of performance variance (Anthropic, 2025). Agentic token consumption is the dominant line item, not a rounding error.
Why do longer agent loops produce worse results?#
Because of context rot. Anthropic's context-engineering guidance describes context as a finite attention budget where recall degrades as the window fills (Anthropic, 2025). Each additional pass works from a noisier window, so output fidelity drops even as cost climbs. Longer loops are simply more expensive and less accurate.
What is a runaway agent cost?#
A runaway agent cost is an unbounded loop that keeps consuming tokens past the point of positive token yield. The agent keeps spending past that line, driving AI agent token cost with no yield to show for it. It is the agent loop cost equivalent of running a machine after it has stopped producing anything worth shipping.
How do you reduce AI agent token cost without losing capability?#
Bound the loop and feed each pass curated, decision-grade context. You do not cut capability by cutting waste. In our controlled test, curated context cut tokens 42% and tool calls 64% on an identical model and prompt (Unblocked, 2026). Fewer, higher-yield passes beat many noisy ones.
What is the difference between agentic token consumption and runaway agent cost?#
The difference is yield, not volume. Healthy agentic token consumption produces decision-grade output per pass. Runaway cost keeps spending past that line. The DORA 2025 report (DORA, 2025) frames the stakes well: AI raises throughput but is negatively associated with delivery stability, because "AI's primary role is as an amplifier."
That phrase is the whole point. A bounded, well-fed loop amplifies into faster, stable delivery. A sprawling, self-running loop amplifies into churn and rework. Same tool, opposite outcomes, decided by the discipline around it.
The field data backs this up. Faros AI (Faros AI, 2025) found high-AI teams merged 98% more pull requests, yet review time rose 91% and there was no org-level delivery gain. More volume, more loops, no yield: that is the runaway pattern in miniature.
Faros AI found that high-AI-adoption teams merged 98% more pull requests while review time climbed 91%, with no measurable org-level delivery improvement (Faros AI, 2025). Higher output volume without higher yield mirrors the runaway-loop trap at the team level.
So the boundary between the two is not a token cap you set arbitrarily. It is a yield threshold. Keep iterating while each pass clears the bar for decision-grade output. Stop when it does not. This is the cost cousin of the quality failure we documented in the AI agent doom loop, where unbounded autonomy degrades the work itself. The problem is volume without yield.
DORA's 2025 research finds that AI increases throughput while showing a negative relationship with delivery stability, concluding that "AI's primary role is as an amplifier" (DORA, 2025). Whether a loop amplifies value or waste depends on yield discipline, not token volume.
How do you bound an agent loop so every pass earns its tokens?#
You bound a loop by pairing limited iterations with curated context and a yield-based stop condition. The business case is stark. McKinsey's State of AI 2025 (McKinsey, 2025) found only 39% of organizations report any EBIT impact from AI. Autonomy without yield discipline does not convert to results.
Start with bounded iterations and a stop condition tied to output quality, not a fixed pass count. Then make every pass start from the full picture rather than a degrading window. That second part is where context engineering does the heavy lifting, and it is where loops earn their tokens instead of burning them.
This is the inverse of the routing failure we cover in model routing for coding agents, where teams downgrade to cheaper models that need more retries. Both failures ignore yield. As our customers put it:
"You cannot make coding agents work without domain and functional context... an agent gets the full picture, not just code analysis, but why decisions were made and what the constraints are."
— Raphael Bres, CTO, Tradeshift
Unblocked is the context engine for engineering, so bounded loops beat unbounded ones. For the operational side of bounding autonomy, see stop babysitting your agents.
McKinsey's State of AI 2025 reports that only 39% of organizations see any EBIT impact from their AI use (McKinsey, 2025). Autonomy alone does not move the bottom line. Bounded loops fed decision-grade context, where every pass earns its tokens, are what convert agent activity into measurable business yield.
Loops Worth Running#
The question was never how long the model runs. It is whether each pass is worth running. Self-running loops sold as efficiency multiply AI agent token cost while context rot drags fidelity down, so you pay more for less. Bounded loops fed curated context invert that trade. Fewer passes, higher yield, every pass starting with the full picture.
Anthropic's own numbers settle the debate: roughly 15x the tokens for multi-agent systems, with token usage explaining about 80% of performance variance. That is not an argument for running longer. It is an argument for running smarter, with the full token yield framework as your scorecard. Bound the loop, feed it well, and measure yield, not runtime. The loops worth running are the ones that earn their tokens.
