Chatting with agents

An agent answers from its instructions, the conversation, and whatever you attach. Agents with web search enabled can also search the open web. Sonata can also connect agents to systems your organization runs, like Google Workspace, under your governance; when an agent uses a connected tool, a quiet line in the conversation names the call, like Google Drive: search files.

When an answer draws on the web or your documents, its sources are listed with the answer, each a real link. Citations carry into the Word export as footnotes. An answer that rests on sources is one you can check, and Sonata expects you to: treat agent output as a strong draft for your review, not a finished opinion.

Sonata draws a hard line between reading and acting. By design, reading is open, but any action that would change something outside Sonata, like sending an email or creating a file, is built to pause the conversation and show you exactly what the agent wants to do. Nothing happens until you approve that specific action; decline it, and the agent carries on without it. This is not a setting you can turn off, and that is deliberate.

If your organization uses Knowledge collections and in-chat research is available, an agent can research them mid-conversation for small questions, reading up to 15 documents and citing what it finds. For bigger questions, the agent points you to the Research page, which is built for corpus-scale work and always available.

Seven agents do their core work in code before the model says a word, and each has its own guide: Comparing two documents, Comparing bill versions, Compiling a bill’s legislative history, Digesting a congressional hearing, Analyzing a docket’s public comments, and Mapping a lobbying landscape.