**AI agents still diverge at temperature 0. When does it matter?**

> AI Agents Still Diverge at Temperature 0. When Does It Matter? 12 of 13 runs reached the exact same code state, then ended 11 different ways — even at temperature 0. What should we do about … AI AI Agent Artificial Intelligence LLM Machine Learning AI Agents Still Diverge at Temperature 0. When Does It Matter? Ruxi Zhang 9 min read · 10 hours ago -- Listen Share 12 of 13 runs reached the exact same code state, then ended 11 different ways — even at temperature 0. What should we do about it? Abstract I ran the same coding agent on the same task repeatedly, using the same model, configuration, and temperature=0 . Twelve of thirteen executions reached the exact same intermediate repository state. Those twelve runs subsequently ended in eleven distinct source states. At first, I thought the problem was variation itself. I no longer think that is the right framing. Different executions may still produce equally correct solutions. The more useful question is whether we can identify when harmless exploration becomes consequential divergence — and intervene before the agent fails. Same state. Different futures . Press enter or click to view image in full size 1. Observation: Temperature Zero Does Not Make an Agent Execution Deterministic An agent is not a single model invocation. It is a closed-loop system in which model outputs produce actions, actions change an environment, environmental changes produce new observations, and those observations become part of subsequent model context. Even a small difference early in this loop can therefore alter later tool calls and observations and eventually produce a substantially different execution. To study this behavior, I repeatedly ran the same coding agent on SWE-bench tasks using the same model and temperature=0 . Rather than interpreting the model&\#x27;s reasoning text, AgentSeism records an objective environment-level representation after each action. For coding tasks, a source state is the canonical tracked Git diff of the repository against the task&\#x27;s base commit. Scratch files and other untracked workspace artifacts are recorded separately. This representation revealed three recurring trajectory topologies. In absorption , two executions diverge but later reconverge and finish in the same source state. In persistence , they diverge and remain different through termination. In re-divergence , they diverge, later reach exactly the same non-empty source state, and subsequently separate again. Three observed fates of execution variation Press enter or click to view image in full size The important observation is not simply that executions differ. In one Pytest task, independent runs repeatedly reached the exact same non-empty source state and subsequently produced different final source states. Across the executions summarized in Figure 1, 12 of 13 runs reached that shared intermediate state, yet those runs ended in 11 distinct final states. This suggests that repository-source convergence does not imply convergence of the complete agent state, because different executions can arrive with different message histories, observations, and scratch workspace artifacts. 2. Reframing the Problem: Variation Is Not Failure My initial framing treated persistent or re-divergent trajectories as evidence of instability. That interpretation is too strong. Two executions may follow different paths and produce different patches while both satisfy the task specification. Conversely, an agent can behave very consistently and repeatedly produce the same incorrect answer. Behavioral reproducibility and task reliability are therefore different quantities. For coding agents, this distinction can be measured without relying on ano...

**来源信息**
- **来源**：Hacker News：AI 热帖
- **分类**：论文
- **发布时间**：2026-09-12 00:34（北京时间）
- **原文**：[打开原文](https://medium.com/@ruxiz2005/ai-agents-still-diverge-at-temperature-0-when-does-it-matter-f5abac008e17)