GLM-4.7 Released: Preserving Reasoning State Across Tool Steps for Long-Chain Coding Agents
Zhipu AI released and open-sourced GLM-4.7 weights in December 2025, focusing on enhanced coding, tool use, and multi-step reasoning.
GLM-4.7 pushes the competitive frontier to state preservation during multi-turn agent execution: the model needs to reuse existing reasoning across consecutive tool calls rather than restarting each step.
The official introduction describes Interleaved Thinking, Preserved Thinking, and Turn-level Thinking, enabling the model to think before tool calls, retain thinking blocks across turns, and control reasoning budget per turn.
The bottleneck for coding agents has shifted from single-generation quality to long-chain consistency, state compression, and failure recovery; model interfaces are also beginning to expose finer-grained reasoning control.
Engineering teams should decompose tasks into auditable stages and check whether preserved state reduces redundant work, while avoiding carrying untrusted intermediate reasoning into subsequent steps over long periods.
Observe the real impact of preserved reasoning on token cost, error accumulation, privacy leakage, and long-task success rate.