Agent inference costs are exploding as multi-agent workflows burn through tokens

机器之心 · wechat · 2026-07-26

AI agents are moving from demos into enterprise workflows, and their biggest pain point is turning out to be inference cost.

Core argument

Agent systems consume far more tokens than ordinary chat because they run in repeated think → act loops.
Long-running, stateful agents can amplify even small failures, while multi-agent setups add overhead for task splitting, synchronization, and aggregation.
The article cites Anthropic research showing a single agent can use about 4× the tokens of a normal one-shot chat, while a multi-agent setup can exceed 15×.

Why costs blow up

In ReAct workflows, each cycle re-encodes a growing history, so token usage compounds with every step.
Plan-and-Solve architectures can make this worse by splitting work into many sub-steps, each of which may trigger its own ReAct loop.
In extreme cases, agents can get stuck in loops and run far longer than expected, making spend hard to cap in advance.

Cost-control strategies the article highlights

Put adaptive reasoning budgets into the model itself.
Add request-level routing between the agent layer and the model.
Filter redundant context before sending prompts.
Compress fixed prompt overhead and reduce repetitive execution paths.

Real-world examples

A MiHoYo engineer reportedly left dozens of agents running over a weekend and came back to a loop that burned roughly RMB 2 million in tokens over 13 hours.
OpenClaw founder Peter Steinberger shared a bill showing a 3-person team using 100 Codex agents, with 603 billion tokens, 7.6 million requests, and a $1.3 million monthly bill.

Bigger question

If frontier models start baking in adaptive thinking and cost controls, the article asks whether third-party agent orchestration layers will still have a strong role to play.

Original post →

Agent inference costs are exploding as multi-agent workflows burn through tokens

More from coding & agent