Reading a reply: the translator and expanding turns
An agent’s reply in Lowkey isn’t the raw output of the AI engine. It’s a translation, and the raw work is always one tap away. Knowing this explains most of what you see in a conversation.
The translator
Section titled “The translator”While an agent works, its engine produces a lot of raw material: commands it ran, files it read, notes to itself, long technical output. When the turn finishes, a second model, the translator, reads all of that and writes the reply you see: short, plain, in that agent’s own voice and personality.
That’s why a reply can read differently from what a tool printed. The translator summarizes; it doesn’t change what happened. If you need the exact wording, the exact numbers, or want to check how an answer was reached, expand the turn.
When an agent thinks about a message on several models at once, the translator is also what merges their answers into one reply (see Choosing the model, or several at once).
Expanding a turn
Section titled “Expanding a turn”Every reply has a footer line at the bottom. Tap it to expand that turn and see the raw work underneath: the agent’s own unedited response, each action it took (commands, file edits, searches, workers it started), and timing. Tap again to fold it away.
Nothing is hidden. Workers an agent started appear as cards you can open, and their own conversations show everything they did too.
Expanding every turn
Section titled “Expanding every turn”To open all turns in a conversation at once, use Expand all turns in the conversation’s header (the same button collapses them again). This is handy when you want to follow a long piece of work step by step, or you just like watching the machinery.
If you always want detail, make it the default: Settings → Appearance → Expand turns. New conversations then open with turns expanded. The default is saved on the device you set it on, and any conversation can still be toggled from its header.
Staying in control
Section titled “Staying in control”Expanding or collapsing changes only what you see, never the agent’s work. If a reply ever seems to say something the raw output doesn’t support, tell the agent; the raw response is the record.
Your admin (or the Lowkey team) can adjust how the translator writes (its model and template) in Settings → Translator.
What to read next
Section titled “What to read next”- Who’s doing the work you’re reading about → Talking to agents
- Asking several models at once → Choosing the model, or several at once
- Conversations, turns, stopping and rewinding → Conversations, sessions, and turns