Agents

An agent is a saved assistant with its own name, instructions, model and toolset. Build it once, then talk to it, schedule it, hand it work from chat, or put it on email, Slack, Teams or a public page.

Ordinary chat starts fresh every time. You explain who you are, what you want and how you like it done, and the next conversation forgets. An agent keeps all of that: a job description, the tools it may use, the skills and documents it should lean on, and a history of every conversation it has had.

Think of it as hiring for one role. A Lead Qualifier, a Support Triage desk, a Weekly Report writer. Each one is its own agent with its own settings, chats, insights and channels.

Open Agents from the left sidebar (/user/agents).

Agents are part of the paid plans. Without a plan you can still open the Agents page and see anything shared with you, but Create Agent is disabled and the page shows a Purchase plan banner.

The Agents page

The gallery at /user/agents lists every agent you can use.

Click a card to start chatting with that agent. Its settings are one click further in.

Each card has a ⋮ menu:

Creating an agent

1. On /user/agents, click Create Agent.
2. Volnyn creates a blank agent called New agent and opens its workspace at /user/agents/{id}.
3. Click the pencil next to the name to rename it.
4. In the settings panel on the right, write its instructions under Agent Preferences. Write them the way you would brief a new colleague: the role, the goal, the steps, the tone, what to avoid.
5. Optionally pick a model, attach Skills and Knowledge Sources, and set its Abilities.
6. Click Save, or just send it a message. Unsaved edits show a banner: "Send a message to save & continue with your changes", with an OR revert to original link.

You can keep up to 100 agents and 40 folders.

The agent workspace

The workspace (/user/agents/{id}) has two halves:

• Left: a live chat with the agent. Type in "Send a message to your agent", attach an image, or use the microphone. Replies show their Thought process when the model reasoned first.
• Right: the settings panel, with two tabs, Agent and Settings. Toggle the panel with the arrow button in the top bar; the page remembers whether you left it open.

The top bar also has:

• The agent's name — click it for Chats with this agent, a menu of past conversations.
• New Chat — start a clean thread. If you are already on an empty one, it stays put rather than creating another.
• Copy link to this agent.
• Chat details — opens Details (conversation facts and a Credits breakdown for this chat) and Files (everything the conversation produced).

When you open an agent, the sidebar switches to that agent's own menu: New Chat, Artifacts, Reflections, Evaluations, Insights, Chats, the External Channels, and a searchable Recents list. The arrow at the top takes you back to all agents.

Settings panel: the Agent tab

Agent Preferences

Advanced jumps to the Settings tab.

Triggers

How the agent starts work without you. Click Add to open Automations and give it a schedule or an event.

The AI Managed flag (on by default) lets the agent create automations and scheduled tasks itself when you ask for something recurring, such as "check this every morning". It still asks before creating one: a card appears in the chat, Create this automation? or Create this scheduled task?, and nothing is created until you confirm. Turn the flag off and the agent cannot set up triggers on its own.

Connectors

The apps the agent can reach — Gmail, Slack, Sheets, Notion and the rest. Connected apps show as icons; Add Connector opens chat, where you connect new ones. See App connectors.

AI Discovery (on by default) lets the agent use any app you have connected. Turn it off to keep the agent away from connected apps entirely.

If the agent needs an app you have not connected, it pauses and shows a Sign in with… card in the chat. Sign in in the new tab and the agent picks up where it stopped — you do not need to resend the message.

Skills

Attach saved Skills with Add. Attached skills are the methods this agent follows.

AI Skill Editing (on by default) also lets the agent look up and run any of your other skills when a task calls for one. With it off, the agent only uses the skills you attached.

Knowledge Sources

Attach documents from Brain with Add. The agent searches them when a question is about your own material, and cites what it found.

Subagents

Other agents this one can hand work to. Click Add to pick from your agents.

The first row is the agent itself, marked (Me): "Lets this agent clone itself as a subagent." The ⋮ on that row clones it — the copy opens as its own agent and is added to this one's subagents.

Reflections

"Each night this agent reads its own recent work and proposes what to change about itself." The flag shows the current mode (Review queue or Auto-apply: ON). Click Review to open the Reflections page. See Reflections below.

Abilities

The groups of tools the agent may use. Each row is a switch:

Only abilities your account can actually use are listed.

Leave every ability off and the agent can use everything. Turning some on narrows it to exactly those (in the default Tool Discovery mode).

Tool Discovery (the small badge next to Abilities) cycles between three modes:

If a task needs an ability the agent does not have, it says so in the chat with a Turn on … switch. Turning it on only ever adds that ability; the paused task then continues.

Settings panel: the Settings tab

Talking to an agent

There are three ways in:

• From the Agents page — click the card. It opens the agent's current thread in the main chat, at /user/agents/{id}/c/{conversation}.
• From the workspace — use the chat on the left of /user/agents/{id}.
• From any chat — ask for it by name: "Ask my Researcher agent to find three competitors for…". The assistant finds your saved agents and runs the one you named.

When the agent wants to do something consequential, it asks first. An approval card appears with Confirm and Reject; nothing happens until you choose.

Every conversation with an agent is kept. Find them under Chats in the agent's sidebar (/user/agents/{id}/chats), or in the Recents list.

Agent pages in the sidebar

Reflections

On a schedule you choose, the agent reads its recent conversations and looks for patterns — the same correction three times, a tool it keeps missing, a request it keeps getting wrong. Each finding becomes a proposal card with a rationale and evidence. One-off corrections are not enough; it needs to see a pattern repeat before proposing anything.

On the Reflections page (/user/agents/{id}/reflections):

• Proposals are listed newest first, with a calendar that dots the days that have some.
• Apply an instruction change and its wording is added to the agent's instructions, exactly as shown on the card.
• Some proposals — a new skill, a new tool — are changes you make yourself. Applying those marks them done.
• Decline removes a proposal from the queue.

The page's Settings control:

A run with no new activity since the last one is skipped before anything is spent.

Insights

/user/agents/{id}/insights shows the agent's activity over a window you pick: chats and messages per day, tool calls and which tools, credits per day and per chat, evaluation pass rate and the rules that failed, and how many reflections were proposed, applied or declined. Ask about this sends the numbers to the agent's chat as a question you can edit before sending.

External channels

Under External Channels in the agent's sidebar, each channel has its own setup page at /user/agents/{id}/channels/{channel}. Every channel has the same core controls:

The page also shows how many messages the channel has handled and when it was last used.

If a channel is not ready — Slack not connected, inbound email not configured on the server — the page says exactly what is missing.

Sharing with your team

When you are on a Team, use Share or Move to Team on the card. Teammates then see the agent under Shared with me and can chat with it. Only the owner and team editors can change its settings; viewers cannot run agents.

Deleting an agent

Use Delete on the card's menu, or Delete agent on the workspace's Settings tab.

• The agent and all of its chat history are removed. This cannot be undone.
• Its scheduled tasks and automations are switched off, not deleted. Review them on the Automations page and decide whether to keep them.
• Deleting a folder never deletes the agents inside it; they simply lose their folder.

You can always delete agents you own, even if your plan has lapsed.

Tips

• Write instructions as a brief, not a prompt. Role, goal, steps, tone, and what never to do.
• Start broad, then narrow. Leave Abilities off while you try the agent out, then switch on only what it actually used (Insights shows which tools it called).
• Attach, don't paste. Put reference material in Brain and attach it as a Knowledge Source instead of pasting it into the instructions.
• Let Reflections do the tuning. Keep the default review queue and accept the proposals that match what you keep correcting.
• Lock down public channels. Set Only answer these senders on email, and keep the hosted page link private unless you want the world billing your credits.

FAQ

What is the difference between an agent and a skill?

A skill is one method written down — instructions for a single job. An agent is a whole assistant: instructions plus a model, tools, skills, knowledge, subagents, a chat history, triggers and channels. Agents use skills.

Does an agent remember past conversations?

Each chat is its own thread, and all of them are kept under Chats. Turn on Search Past Conversations to let the agent look back through earlier chats.

Can an agent act without me?

Yes, through Automations (Triggers) and External Channels. It still asks for approval before creating new automations or scheduled tasks.

Why does the agent say it needs to sign in?

It needs an app you have not connected. Click Sign in with… on the card; the agent continues on its own once you have.

Why can't I create an agent?

Creating and running agents requires a paid plan. The Agents page shows a Purchase plan banner when your account does not have one.

What does running an agent cost?

Agent runs use credits like any other chat. Messages that arrive through email, Slack, Teams or the hosted page are billed to the agent owner's credits. See the Credits section of Chat details for a per-chat breakdown, and Insights for totals.

Next steps

• Create your first agent at /user/agents and give it a one-paragraph brief.
• Attach a Brain document and a Skill to it.
• Give it a schedule from Automations.
• Let it browse the web for you with the live browser.