How to save Agent's Token bills?

This article is part of the PRO Members Communication, and follows "PRO Members of Machine Heart" at the end of the article to read more on the topic.

The AI industry is undergoing a structural leap from generative capability to Agent agility, whose core contradictions have shifted from the boundaries of model capability to the controllability of reasoning costs. As industry leaders such as finance, manufacturing, and the Internet have incorporated Agent into automated flowlines, Token consumption has received widespread attention. Since 2026, from model-side adaptive reasoning mechanisms to routing-side intelligent scheduling, the industry's attention has begun to focus on "downstream" technology routes.

Directory

01. What led to the escalating cost of reasoning for Agent?

What is the mechanism of anthropic self-adaptation thinking to reduce costs from the model side?...

What caused Agent's reasoning costs to climb the ladder?

1. In the past few months, AI has accelerated its move from the lab to enterprise production, with enterprises using Harness to access Agent to automation processes. At the same time, Agent's Token consumption has increased exponentially, model manufacturers want enterprises to expand their use, enterprises want to control costs, and demands between the two are pushing the industry to find solutions to technology reductions.[1-1]

1 The state-of-the-art characteristics of a smartphone require it to perpetuate execution code and process errors during long-term operation, and small-scale system failures can have a disastrous impact on the smartphone in the absence of effective mitigation measures, leading to a sharp expansion in token consumption.[1-2]

2 Multi-intelligent collaboration systems further amplify this cost pressure. In addition to single-agent cycle consumption, task allocation, state synchronization, and result aggregation overlap additional costs. Research data from the Anthropic team shows that Token consumption for running a single Agent completing a standard task is about four times that of traditional single-cycle conversations, while Token consumption under the multi-intelligent collaboration architecture reaches more than 15 times that of standard conversations.[1-2]

2 Compared to the single-cycle request-response model of traditional big language models, Agent's ReAct reasoning paradigm breaks down tasks into multiple "think-action" cycles, each of which requires re-encoding the entire historical context. Token consumption continues to accumulate with the depth of the cycle, and cycle termination conditions depend on the complexity of the task and are difficult to predict beforehand.[1-3]

1 In complex tasks, the agent may need dozens of rounds of reasoning-action cycles to get the answer, and in marginal cases, the agent may even fall into a cycle that cannot end on its own, resulting in an infinite increase in Token consumption.[1-3]

3 Even if the cost of a single cycle is still acceptable, the cumulative increase in the above context window constitutes another layer of cost enlargement when the Agent breaks down the task into several or even dozens of execution steps and triggers a ReAct-like multi-cycle cycle independently at each step.[1-4]

1 Plan-and-Solve is an instructional architecture that breaks down tasks into two phases of "preplanning, then executing", with the model master as an orderly multi-step execution plan, followed by a step-by-step execution of each subtask.[1-4]

4. The current popular multi-smart body workflow design has led to a significant increase in reasoning costs, which has continued to increase the industry's focus on Agent reasoning costs, further adding to the cost pressure on enterprises in the event that they fail to produce effective results after burning millions of tokens.

1 In 2026, an engineer deployed dozens of Agents for multi-intelligent collaboration tests over the weekend, and without setting a Token consumption ceiling, the Intelligent enters a loop call, running continuously for 13 hours, burning a large model Token worth 2 million yuan.[1-5]

2 OpenClaw founder Peter Steinberger unveiled his OpenAI account, with a three-person team commanding 100 Codex smartphones, consuming 60.3 billion tokens in 30 days, sending 7.6 million requests, and accounting for up to $1.3 million (approximately 93.3 million yuan).[1-6]

What reduction schemes have been designed by the industry to reduce the cost of Agent reasoning?

1. In the increasing trend of smart-body reasoning costs, the industry is exploring reduction schemes from multiple dimensions. For example, model manufacturers will reason budget control built into the model itself, some open-source projects will add request-level routing between the Agent and the model, and other tools will filter the context from the data input side to filter redundancy or compress fixed consumption from the tip-side.