Agent failures and model selection issues across all regions

Update timeline

1. investigating Oct 01, 2026, 06:57 PM UTC

   Starting October 1 at approximately 1:30 PM UTC, some Agents, including both low-code and coded agents, began experiencing failures across all regions when using gemini-3.1-pro-preview and gpt-4o-2024-05-13. These models may also be unavailable for selection during agent design. We are actively investigating and working on a fix.

2. identified Oct 01, 2026, 07:05 PM UTC

   We have identified the issue and are deploying a fix. Agents using the affected models may continue to experience failures, and the models may remain unavailable for selection during agent design, while the deployment completes.

3. resolved Oct 01, 2026, 07:46 PM UTC

   We have reverted the change that caused this issue and confirmed that the affected models are available again. Some agents may continue to experience intermittent failures while instances refresh. We are continuing to monitor agent design and execution for gemini-3.1-pro-preview and gpt-4o-2024-05-13.

4. postmortem Oct 07, 2026, 10:15 AM UTC

   ## Customer impact On October 1, 2026, from 14:31 to 19:40 UTC, agents configured with some large language models, including Gemini 3.1 Pro Preview and GPT-4o (2024-05-13), failed to start in all regions, with an error saying the model is not available. The same models were missing from the model list while designing agents. Agents using other models were not affected. ## Root cause A feature flag was flipped, causing Agents service to begin sending an operation code to model discovery requests. This caused the list of models served to be that specific operation’s as opposed to the base agents model list. The per operation model list was initially seeded from the Bring Your Own Model (BYOM) configuration of overridable models, which crucially for agents did not have parity with the full list of supported models in Agents service. Internal audit showed this was an error and not an intentional gap. Over time this gap was never addressed and after the flag switch, agents already configured with any of the affected models would fail because of their absence from the returned list. ## Detection A customer reported the issue, and it was escalated to our engineers at 16:44 UTC. Our monitoring did not alert, because this error usually points to organization or agent level configuration issues. ## Response Engineers traced the failures to the configuration change and reverted it at 19:01 UTC. The revert took effect over the following minutes, and the last failure occurred at 19:40 UTC. ## Follow up 1. The model lists will be aligned before this change is made again. 2. We're changing how this error is classified so a platform-wide spike alerts our engineers.