Unverified50% confidenceFactExact time
Agents repeatedly execute the perception-planning-action loop in cycles, and each round requires packing the complete history into the context window, causing input tokens to grow linearly or even exponentially.
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Codex超高推理模式隐藏成本：一个任务耗尽70%额度的真相
redditr/OpenAI7/12/2026
Related Claims
UnverifiedAgentExecutor每次工具调用前会重新拼接完整对话历史进Prompt，导致Token消耗随任务步骤线性增长78% similarUnverifiedEach iteration cycle of an Agent consists of three phases—model inference, tool invocation, and result return—and latency accumulates at every step77% similarUnverifiedIn Agent scenarios, each tool call requires the full context to be re-input into the model, causing token consumption to grow exponentially.77% similarUnverifiedIntelligent behaviors in Agents such as planning, reflection, and research fundamentally depend on precisely assembling the environmental state, historical memory, and task objectives into context at each moment before inputting them into the model.76% similarVerifiedThe Agent Loop is essentially a While loop that continuously thinks, acts, rethinks, and acts again until the goal is achieved.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/608400API
curl https://kongchang.com/api/v1/knowledge/claims/608400MCP
get_claim(id=608400)