Skip to main content
An agent is a named worker for one job, such as inbound leads or outbound prospecting. It sits above your automations: in the app, triggers are called automations, and every automation belongs to exactly one agent, which supplies its goal, tools, memory, and scope. Each agent has:
  • A name and a goal. The goal says what the agent is for and why (“Find net-new accounts that fit the ICP and start conversations with 30 of them every month”). It explains the purpose behind the agent’s automations but never authorizes an action on its own: each automation’s instructions decide what a run does. A goal can be up to 1,000 characters.
  • Instructions. Optional standing rules every run of the agent follows, whichever automation fired: who to route to, what never to do, which CRM fields to set. Where an automation’s own instructions are more specific, they win. Instructions can be up to 3,000 characters.
  • Tools. The toolkits and connections its automations can use (see below).
  • Picked skills. Skills listed first in its runs, so Swan reaches for them first. Every other skill stays available.
  • One memory. Notes its runs keep for later runs (see below).
  • Its reminders. One-time follow-ups set by its runs or by anyone who can edit it; each runs as the agent when it’s due (see reminders).
  • Its automations. Every automation belongs to exactly one agent.
“Agent” here means one of these workers. It isn’t Swan itself (the assistant you chat with), a sub-agent (a run Swan starts for one task in a batch), or an external agent connected to Swan over MCP. Agents are under Agents in the app’s sidebar, and you can also ask Swan in chat to create, change, or explain one. New agent opens a chat with Swan, which asks what the agent is for and sets it up.

The agent page

Each agent has its own page, with up to two tabs:
  • Overview — the agent’s status, owner or creator, triggers, and integrations, then its Runs, newest first. Open a row to see what the run did, or open the run for its full page.
  • Needs you — the decisions the agent’s runs are waiting on from you.
Beside Runs are the agent’s goal, instructions, triggers, tools, and skills. If you can edit the agent (see below), change its name and goal in place, its instructions on their own page, and its tools and skills from their panels. Manage on the Triggers panel opens the agent’s triggers, where you turn each one on or off and open it to edit. To add a trigger, ask Swan in chat. A trigger, a run, and a ticket each open inside the agent’s page. Outreach sequences appear on the run that created them.

Org and personal agents

  • Org agents are created and edited by Admins and visible to everyone in the org. Org automations live in org agents, so only Admins create or edit them. Members can read org agents, including their memory, but can’t change them.
  • Personal agents are created by Members and Admins for themselves; Viewers can’t create agents. A personal agent is visible only to its owner; Admins can’t see it either, though they still see its runs. Your personal automations live in your personal agents.
Creating an automation always needs an agent: pick an existing one, or create one on the spot.

Tools belong to the agent

Every automation run has Swan’s built-in tools, whatever its agent: files and code execution, records and companies, skills and tags, research and prospecting, sequences, tickets, presentations and file sharing, run history, and Slack when it’s connected. The agent’s tools decide which connections its runs can use:
  • connected apps and MCP servers;
  • HubSpot, Email, LinkedIn and the browser, which are off unless the agent lists them.
There’s no tool picker on an individual automation. Adding a tool to an agent gives it to all of that agent’s automations, and removing one takes it from all of them. When you change an agent’s tools or move an automation, Swan checks the affected automations’ instructions and warns you if one needs a tool its agent no longer has. Every agent can use org connections, and a personal agent can also use its owner’s personal connections, such as their Email and LinkedIn inboxes. An Admin can give an org agent a member’s Email or LinkedIn; every automation in that agent can then act through it, so keep automations that need it in an agent of their own. Chat can’t use it unless it’s available through an org agent (see asking an agent from chat), and Members and Viewers can’t add another member’s connection. Personal connections are for direct Email or LinkedIn actions performed by the run; inboxes used by outreach sequences are selected separately (see inboxes).

Agent memory

Each agent keeps one memory: what its runs have learned, such as what worked, what to avoid, and where the last run left off. The whole memory is loaded into every run of the agent’s automations, so each run builds on earlier ones instead of starting from scratch. Only that agent’s runs write to it, along with the people who can edit the agent (by asking Swan in chat). An org agent’s memory is visible to the whole org. Like record memory, it’s capped at 10,000 characters. See agent memory.

Agent heartbeat

Every few days (three by default, reviewEveryDays in AGENT.md, 1 to 30), each agent checks how far it is from its goal and moves it closer. It runs on any day of the week, in the working hours of the person who gets its tickets. An org’s agents are reviewed on different days. On its own, it fixes, rewrites, retunes or pauses its automations, and sets how often it reviews: sooner while things fail, later when the agent is on track. Ask Swan for a specific cadence and it writes it into the agent’s instructions, which the heartbeat keeps. It follows the agent’s instructions: what they forbid, it doesn’t do. Each change is logged with the old and new text in the agent’s HEARTBEAT.md, so you can undo it, and your daily heartbeat tells you about it. It asks first, on one ticket with the exact change to approve, for:
  • Any change to the agent’s AGENT.md (instructions, tools, skills) or to a skill.
  • Any change, other than pausing it, to an automation that sends outreach without a person approving each sequence.
  • A change that makes an automation start doing outreach, or start sending without a person’s approval.
  • A change that can raise credit use: running more often, a new or re-enabled automation, added enrichment or scraping, or bigger batches.
  • A change that makes an automation delete records, files or memory.
  • Whatever only you can do.
It never changes the goal and can’t turn itself off. A personal agent’s tickets go to its owner. An org agent’s go to an Admin: by default whoever created the agent, or, when that person is not an Admin, the Admin who created most of its automations, or else the longest-standing Admin. Name someone else in the agent’s instructions to change that.

Moving automations and agents

  • An automation moves only between agents of the same scope: org to org, or personal to personal. From then on it uses its new agent’s tools, goal, and memory.
  • To change scope, move the agent. Its automations and memory go with it.
    • Personal to org: only the agent’s owner can do this, and only if they’re an Admin. Its memory becomes visible to everyone in the org, and Swan asks you to confirm first.
    • Org to personal: only an Admin can do this, and the agent becomes theirs.
    • The move is refused if any of the agent’s automations can’t exist at the new scope. Website visitors, business events, Bombora intent, and Website Changes automations are org-only. Some app event automations also work at only one level. Moving to the org is also refused while its skills include one of its owner’s personal skills, and moving to personal while its tools include another member’s Email or LinkedIn. Either way, the move is refused if the other scope already has an agent with the same name; rename one of them first.

Asking an agent from chat

In chat you still talk to Swan, not to an agent, but Swan can hand work to one. When you ask for something that is an agent’s job, such as “have the outbound agent follow up with these 20 accounts”, Swan delegates each item as a sub-agent that works as that agent: it starts with the agent’s goal, instructions, picked skills and memory, uses the agent’s tools, and adds what it learns to the agent’s memory, like a run of one of its automations. It can’t change agents or automations, or set reminders. Swan can hand work to any org agent and to your own personal agents. That includes an org agent that has a teammate’s Email or LinkedIn, so its sub-agents can act through that inbox even when a Member asked. Only chat can do this: an automation’s sub-agents always work as its own agent, never as another one.

Deleting an agent

Deleting an agent also deletes its automations, its memory, its files and its reminders, and the confirmation shows how many automations go with it. Deleting one of its automations doesn’t cancel the agent’s reminders. When a member is removed from the org, their personal agents and the automations in them are deleted too, and their pending reminders don’t fire.

Common questions

I had automations before agents. What happened to them? When agents launched, Swan put every existing automation, paused ones included, into an agent for its kind of work, such as Signal-led selling, Inbound or RevOps. One-time reminders set before then were left out (they later became reminders of the person who set them), and so were the Message Received automations, because replies now go back to whoever started the outreach (see Message received). Swan also deleted two automations it had set up itself: the “Daily check-in” from onboarding, which your daily heartbeat replaces, and the “Trial health check”. Org automations went into org agents and personal automations into their owner’s personal agents, so a personal agent can have the same name as an org agent that does the same kind of work. Each agent started with a goal and a short description of its job written from what its automations do, the combined tools of its automations, and the skills their instructions referred to. Because an agent holds the tools of all its automations, an automation may now be able to use a tool it did not have before, including a teammate’s Email, LinkedIn or calendar. The file migration/MIGRATION.md in your Brain lists every agent and every tool an automation gained, so you can ask Swan what changed. Review your agents: rename them, rewrite each goal in your own words, move automations between agents, and remove tools an agent doesn’t need. Do runs of the same automation remember what earlier runs did? Yes, through their agent’s memory: every run of the agent’s automations starts with it, and runs add what they learn. Automations whose instructions told them to rewrite their own instructions can’t do that anymore; they keep notes in agent memory instead. Does an agent check its own automations? Yes, every three days by default. See agent heartbeat. Can one automation have different tools from the rest of its agent? No. Tools belong to the agent, and every automation in it gets all of them, on top of Swan’s built-in tools. Move the automation to an agent with the right tools, or create a new agent for it. Can an Admin give an org agent a teammate’s Email or LinkedIn? Yes, by adding it to the agent’s tools; every automation in that agent can then use it, so keep automations that need it in an agent of their own. How do I move an automation to another agent? Ask Swan to move it. An automation moves only between agents of the same scope. To turn personal automations into org automations (or the reverse), move their agent; see moving automations and agents. Can I ask a specific agent to do something? Yes, ask Swan in chat. Swan hands the work to that agent, and each task runs as the agent with its tools and memory. See asking an agent from chat. What happens to an agent’s automations when I delete it? They’re deleted with it. The confirmation shows how many automations will go. If you want to keep some, move them to another agent of the same scope first.