AI agents: loops cut prompting but spike tokens 25x

Anthropic's Claude Code team published a term last week that captures the next technical divide in enterprise AI: the agentic loop. In a company blog, the team described it as self-directed cycles of work repeated until a stop condition is met. The contrast with classic prompting is sharp. Every prompt a human sends starts a manual loop – you type, the model responds, you type again. A loop automates that middle step.

Anand Jain, co-founder and chief marketing officer at CleverTap, put a number on the difference. “A loop is an automated system where the AI prompts itself, or where one agent prompts another, working through a task on its own and checking each step against a goal, without a person typing the next instruction in between,” he said. A single prompt can't know whether its output was good. It cannot retry, tell when it's done, or hold itself to a standard unless a person checks every step. Loops close that gap by giving the system something to check against – a passing test, a validated build, a metric moving the right direction.

The naive read says loops look expensive. Every extra cycle of reasoning, tool use, or verification burns tokens. Srinivas Padmanabhuni, chief technology officer of AiEnsured, said token consumption jumps 5 to 25 times or more. That forces providers to build new infrastructure: prompt caching, state compression, more efficient memory. The cost line moves up.

That framing misses the point. Jaspreet Bindra, co-founder and chief executive of AI & Beyond, argued that the objective is not to minimize tokens but to maximize outcomes. A looping system that produces significantly better results with fewer human interventions can deliver productivity gains that dwarf the incremental compute bill. “If an AI agent is asked to conduct research, compare options, draft a report and fact-check its findings, it cannot accomplish all of that through one prompt alone,” he said. “It needs the ability to review its progress, retrieve additional information, correct mistakes and iterate. Loops help AI systems move beyond prediction and toward execution.”

Prompting is not going away. Bindra said the prompt will become the starting point, not the whole interaction. “Loops will take over the execution layer,” he said.

For an investor building a watchlist, the loop debate distills to one question: who captures the value of those extra tokens? Cloud providers that sell compute by the token – Amazon Web Services, Microsoft Azure, Google Cloud – benefit when usage rises per task. Chip makers that design silicon for iterative workloads gain as well. Software companies that build agents with efficient caching and compression create a moat around cost.

The shift also changes what a good AI product looks like. An employee's job moves from managing every step to defining goals, providing oversight, and making final decisions, Padmanabhuni said. “This evolution transforms humans from prompt writers into system architects.”

Bindra said the loop debate will settle on a simple benchmark. “What happens after the prompt is changing. Increasingly, the prompt will serve as the starting point rather than the entire interaction,” he said. “Loops will take over the execution layer, allowing AI systems to reason through tasks, validate outputs and adapt as needed. The future is prompting combined with loops, where each plays a complementary role.”