Sep 16, 2026
Go Agent V2: Pick the Model, Check the Quote, Replay the Run
Chats run on Go Agent V2. You choose the model and how hard it thinks, an answer shows the sentence it came from, and a run that hits a bad step can be replayed instead of started again.
V2 is the second version of what runs a chat, and the difference shows up in long work. It holds a conversation together across many turns and tool calls, keeps the files you gave it readable the whole way through, and records enough of what it did that the run can be read back afterwards.
Choosing the model
Curated models, and how hard they think: Chat carries a model picker and a reasoning setting, so a quick lookup and a long piece of analysis do not have to run the same way. A run uses the model you chose rather than one it inherited.
The settings name the model: The settings tooltip names the concrete model instead of a generic label, so you can tell what a chat is actually running.
A wider list to choose from: GPT-6 Astra, Gemini 3.7 Flash and Gemini 3.8 Flash are all selectable, and Gemini 2.5 Pro has been retired.
Answers carry their source
Quoted, not tagged: A source is attributed by the text it quotes rather than by a citation marker, so checking where an answer came from is a matter of reading it.
Claims in a case are grounded: A response in a case carries references to the documents behind each claim, rather than asking you to take the summary on trust.
Long runs hold together
Your files stay readable: Documents in a conversation remain available as a long chat is compacted, so something attached early is still there late.
The transcript is complete: Every turn is kept, and every tool call is stored alongside the result it returned, so an audit or a playback shows the whole run.
A malformed step is replayable: A run that hits a malformed tool call can be replayed and debugged rather than failing permanently.
A slow tool ends: A tool result that takes too long returns an error instead of leaving the run hanging.
Late work cannot overwrite a finished answer: A background job that arrives after the fact no longer replaces an answer you already have.
It can see your Workflows
Live results: Workflow results can be read while the workflow is still running, rather than only once it finishes.
Rows and fields are navigable: A workflow row or field opens from the chat, so checking a value does not mean leaving the conversation.
Result files arrive in the chat: Files a run produces are attached where you asked for them, and stay attached when a workflow finishes early.
Bounded, and attributable
Runtime limits: Soft and hard limits stop a run from going on indefinitely.
Tool calls count as activity: An MCP tool invocation is recorded as user activity, with the origin and the sign-in route behind it, so a review can attribute it to the person behind it.
Availability
Go Agent V2 is what chats run on. There is nothing to switch on.
