Text AI Improvements/Updates

Several changes to Text AI Agents are live today:
Improved instruction adherence
Better use of custom follow-up tool
Non-response reasoning surfaced in logs
Model deprecations
Quick Notes on non-response and deprecations:
Non-response
Agents will sometimes choose not to respond based on instructions you configure. This is generally not an approach we recommend configuring towards. The recommended approach is equipping your agent with a conditional tool
 that triggers on a specific condition being met and takes a Turn off AI
 action and can appropriately alert the correct team member through other actions you configure on it. However sometimes, you might configure instructions that lead the agent to believe it is not supposed to respond in a certain situation. This capability was not even present on older AI models, but newer models (gpt 5.4 +, grok 4.5) can do this (and the newest ones, gpt 5.6 series, are more likely to when instructions lead that way).
To account for this, we've given agents the ability to explain why they chose not to respond. This reasoning will show up in the Text AI Logs
 and allow you to adapt your agent's instructions if this is not desired behavior. (You can manually edit your agent's instructions, or copy the agent's reasoning to Ava in the Agent Editor interface and have Ava help you edit it, or simply tell Atlas to check out the contact, and Atlas can guide you through it automatically).
To see logs go to AI Agents
 > AI Text Logs
. A message that was processed, but the agent chose not to respond will show up like this:
There are a few situations where this behavior could
 be beneficial. An example of an intentional use: on a journey scoped agent that takes over at the Appointment Scheduled
 stage, you could insert a rule like "if the user's message is simply acknowledging an appointment confirmation or reminder, do not respond".
Generally speaking
, if you have an agent that is equipped with a goal (e.g. it is tasked with scheduling an appointment), instructing it to not respond is almost always wrong.
 This is because no other downstream actions will take place and you will not be alerted and the conversation will likely resume later via a follow-up. The better approach for these agents is using a conditional tool to wrap things up appropriately in specific conversation ending situations.
Whatever you choose, your logs will now include the reason the agent didn't respond if it didn't.
Model deprecations
All OpenAI models below GPT 5.4 are now deprecated in brayv. If you are pinned to an older model in one or more of your agents, please update to a newer model of your choice, or switch to Lightning Auto
 as the provider
 and Optimized
 as the model
.
This will not break your agents. They will automatically route to GPT 5.6 Terra if they are on a deprecated OpenAI model until you make another explicit choice.