Conversations & sessions¶

Everything you do with an agent happens in a session — a single, saved conversation. This guide covers how sessions behave, why your work survives a closed laptop, and the difference between a full and a read-only session.

Sessions are saved automatically¶

Every conversation is kept and listed in the side rail. Click any past session to reopen it exactly where you left off — the full transcript is there, in order.

Your work doesn't depend on your browser¶

Agents run on our servers. Your browser just watches.

That means:

Closing the tab doesn't stop the agent. A long task keeps running.
Losing your connection doesn't lose your place. When you reconnect, the conversation catches up on everything that happened while you were gone.
You can switch devices. Start on your laptop, check progress on your phone.

Nothing an agent says or does is dropped because you stepped away.

Running several at once¶

The multi-task side rail lets you keep multiple sessions going in parallel. Each one streams its own updates independently, so you can have one agent drafting while another researches.

Background tasks¶

Some agents can spin off background work — a subtask — while your main conversation keeps going. Janet does this when a request breaks down into pieces she can work on in parallel with the main conversation. This is different from delegating to a specialist: a subtask is the same agent working on a piece of your request in the background, not a different agent taking over.

A subtask runs as its own session, so it survives just like any other conversation — closing your tab won't stop it.

When you stop a conversation that still has background tasks running, the Stop button offers a choice: stop just this conversation, or stop it and its background tasks together. If a conversation fails while background tasks are still active, you'll see an alert listing what's still running, with a way to stop the ones you don't need. And if a conversation finishes normally while background tasks are still going, a banner lets you check on them and stop any you don't want to keep running.

Stopping a background task marks it as stopped and keeps it from picking up new work — a step already in progress may still finish before it does. This control only shows up for people with permission to stop sessions; without it, you won't see it for background tasks either.

Read-only sessions¶

A session can be read-only. The agent can look things up, analyze, and answer — but it won't:

remember anything from the conversation,
change its own configuration or code, or
schedule follow-up work.

Read-only is great for exploring ("just poke around, don't remember this") or letting someone trial an agent with no side effects.

You'll see an amber Read-only badge in the session header, a lock icon next to the session in the side rail, and any action that would make a change shown as greyed-out with the option to start a full session instead.

Read-only isn't a permission upgrade or downgrade

Read-only only stops an agent from saving things. It never changes what you're allowed to access. Your account's permissions always apply, in every session type.

Listen mode¶

Need to give a long brief without the agent replying to every line? Turn on listen mode and the agent collects your messages quietly; turn it off and it responds once to the whole thing. You can also toggle it just by saying so. See Working with agents.

Notifications¶

Long tasks don't need babysitting. If a session runs for five minutes or more before finishing, you'll get a notification — an in-app entry and, when your organization has email delivery enabled, you may also get an email from the agent that did the work with a link straight back to the session. Quick back-and-forth conversations don't notify; you're already watching those.

Every notification you've ever received is on Settings → Notifications: newest first, unread ones marked, with a filter for a specific agent and a one-click Mark all read. You can also just ask Janet — "what notifications do I have?" or "mark them read" works in chat. (Other agents can be given the same capability, but Janet has it out of the box.)

Transcripts are permanent¶

Once a message appears in a session — yours, the agent's, a tool result, the opening greeting — it stays. We never quietly remove, hide, reorder, or rewrite past messages. Your memory of a conversation and the agent's always match, which is what makes it safe to say "as I mentioned earlier."

The only thing that shortens a transcript is deleting the whole session — a deliberate admin action that's recorded in the audit log.

Searching your conversations¶

Every account has a search button in the header — click it or press Ctrl/Cmd+K — that opens a search page covering everything you've said and everything your agents have replied, in every session you've been part of, including ones you closed months ago.

Type a few words and matching messages show up with the match highlighted, most relevant first.
Each result shows which session it came from. Click a result to jump straight to that message in its original conversation — no scrolling through the whole transcript to find it.
Search reaches into completed and archived sessions too. If it's in your transcript, it's searchable.

If you're an owner or admin, you'll also see an option to search across every session in your organization, not just your own — useful for checking whether an agent already answered a question for someone else, or reviewing what an agent told a customer.

The quick Ctrl/Cmd+K popup always matches on the words you type, for speed — a search for "cancel" won't catch a message that only says "terminate." For that, open View all results (or go straight to the full search page) and switch the Ranking dropdown from Keyword to Semantic, which matches on meaning instead of exact words, or Hybrid to combine both. Ranking defaults to Keyword; any signed-in user can change it, no admin permission required.

Reviewing and exporting an agent's transcripts¶

Owners and admins get a second way in: open any agent's settings and switch to the Transcripts tab for a search scoped to that one agent, with a toggle between "Everyone who used this agent" and just your own sessions. It's the same conversation search described above, just pre-filtered to a single agent — handy for auditing a shared agent without picking it back out of results across your whole organization.

Need the full result set rather than the top matches? The same search can be exported in bulk as JSON or NDJSON through the API (see the API reference) — useful for handing a record to compliance or feeding it into another tool. Every export is written to the audit log, including one scoped to just your own sessions, since a bulk export is sensitive on its own terms.