Dedicated Agent

A dedicated agent is your own always-on agent — it remembers across chats and keeps working when you're away. Where the default Chat AI assistant answers one conversation at a time, a dedicated agent runs continuously in its own environment: it takes whole tasks and works through them, runs jobs on a timetable, learns your preferences, and keeps everything it has done in one durable conversation.

Dedicated agents are a premium feature. Your plan must include one and an administrator provisions it for your account; until then the agent picker shows Not available on your plan or Not set up yet.

Talking to your agent

There are two places to reach it:

• Chat AI → Agent for this chat. Switch the picker from Default to Dedicated. Each agent keeps its own memory, so switching starts a new chat — the previous conversation stays in your history. Everything else in the composer works as usual: attachments, skills, thinking level, voice.
• The agent console (Open agent console) — the agent's own page: its conversation, with the same composer as Chat AI (files and tables from My Data, skills, the thinking level), a live view of what your agent has been doing (the tools it used, whether they succeeded, how long they took) and its Spend for the period against the alert budget set by your administrator. Charts, Studio pages and files it builds show up in the conversation just as they do in Chat AI.

The conversation is durable: your messages, the agent's replies and anything it produced on its own — a scheduled briefing, a follow-up it decided to do — land in the same feed, whether or not you were watching. Files it produces appear inline (images) or as downloads.

Several conversations at once

Every chat with your agent is its own conversation, and the agent can work on several of them at the same time — up to 10 by default. Within one conversation it answers one message at a time:

• Send a second message while it is still answering in the same chat and you are told your agent is still working on the previous message in this conversation; the answer appears when it is ready.
• If it is already busy in as many conversations as it can handle, the message isn't sent — wait for one to finish and send it again.

Skills and the thinking level are remembered per conversation, so each chat keeps the setup you gave it.

Thinking level

Quick, Balanced and Deep decide how hard the agent thinks and which model it uses: Quick runs on a faster model while the conversation still fits it, Balanced and Deep on the main model with more reasoning. Choose Quick for short questions and quick edits, Deep for analysis you'd otherwise check twice.

Handing over a task (Autopilot)

Give your agent a task — hand over something whole. It works on its own, and checks with you before anything sensitive. Describe the outcome ("reconcile last month's invoices against the bank export and list the mismatches"), press Start the task, and follow the run strip: the plan, each tool it uses, tokens used, and its status — Working, Needs you, Done, Failed or Stopped. Stop ends it early; New task starts another. Tasks can take as long as they need, and run at the thinking level selected in the composer.

If your connection drops while a task runs, the task carries on — the run strip picks it up again when you're back.

Approvals

When a task reaches something sensitive — sending, deleting, paying, changing data outside Finblade — the agent pauses with Your agent needs your permission and waits for one of:

Choice	Effect
Allow once	Just this one action, this time
Allow for this task	This kind of action until the task finishes
Always allow	This kind of action from now on, in every task — permanent
Don't allow	Refuses this action; the agent carries on without it

Scheduled work

The Schedule panel lists what your agent does on its own, on a timetable. There is no form: ask it in chat — "Brief me every morning at 9", "every Friday, summarise the week's tickets" — and the job appears here with its next run, progress (n of total runs or n runs so far) and state: Active, Paused, Running now, Failing or Finished. Per job: Run now, Pause / Resume, Delete (it will stop running and can't be restored; your agent keeps everything it has already sent you).

Agent states

State	What it means
Checking	Confirming whether your agent is available
Setting up	The agent is starting in its own environment (about a minute); it cannot take messages yet
Running	Ready
Paused	Stopped by you (Pause in the console) or an administrator. Its memory is kept; an administrator can start it again
Unreachable	A temporary problem — your work is safe, try again shortly

Do it with the API

Dedicated agent: chat/stream for messages — pick the conversation with chat_session_id, and send reasoning_effort and active_skill when they change (with the ws/dedicated-agent/{agent_id}/ socket for live progress) — messages/ for the default conversation's history, activity/ for the tool log and spend, jobs/ for scheduled work, runs/ for Autopilot tasks and approvals, and kill/ to pause the agent.