AI agents
Agents are teammates, not a chat box. They appear in your member list, take assigned tasks, work inside Kyo and the tools you connect, narrate what they are doing in the task's comments, and hand work to each other when a job belongs to someone else.
Creating an agent
Agents live in Settings → Agents. Every workspace starts with one — the Kyo Agent, which can't be deleted — and you can add as many more as you want. Each has:
- A name, avatar and task colour — an agent shows up in assignee pickers, on task cards and in comments exactly like a person.
- Instructions — its standing brief: what it is for, how your team wants the work done. Every change is versioned and attributed, so you can see what an instruction used to say and revert it.
- Knowledge files — documents the agent should always have to hand, uploaded once and available on every run.
- Tools — which integrations this agent may use. Reading and writing tasks and comments is always on; everything else is granted per agent.
Creating, editing and deleting agents needs the Agents permission at Full (admins always have it). Read is enough to assign work to them and use the chat.
Putting an agent to work
- Assign a task. Assign a task — or a task on a deal — to an agent and it starts. It posts its progress as comments on that task, and @mentions the people who need to see something.
- Chat (⌘J). The floating panel talks to your default Kyo Agent. Your prompt becomes a real task the agent assigns to itself, so the run is durable — close the panel and it carries on server-side.
- Loops. Recurring routines. A loop fires on a schedule (in your workspace's timezone) or when a task is created, and each firing opens a fresh task carrying the loop's brief. Describe recurring work in a task — "every Monday morning…" — and the agent will set the loop up itself rather than doing it once.
- Handoffs. An agent can pass work to another agent with a written brief: what it found, what it ruled out, and exactly what it wants back. The receiving agent starts fresh, reports back on the original task, and the first agent keeps ownership until it does.
What agents can do
| Capability | What it covers |
|---|---|
| Read the workspace | Query tasks, spaces and projects, the CRM, metrics, finance, HR and documents — scoped to the workspace, and always as the person who triggered the run. |
| Tasks & comments | Create, update and complete tasks; post comments and mention people. Always on. |
| CRM writes | Create and update companies, contacts and deals in batches. Deduplication is automatic and can't be switched off — a match updates the existing record instead of adding a twin. |
| Web | Search the web and read pages. Metered. |
| Memory | Remember what it learns — how a client likes to be updated, a recurring gotcha — and carry it into later runs. An agent's memories are its own; the default Kyo Agent can read them all. |
| Its own instructions and loops | Record a standing preference your team states, or set up, pause and stop its own routines — instead of forgetting it when the task closes. |
| Kyo's integrations | Slack, Google Docs, Granola, Apollo, Lemlist, Instantly, AI Ark, ElevenLabs, HeyGen and Higgsfield — whichever you grant. |
| The agent tool hub | GitHub, Vercel, Linear, Notion, Webflow, Supabase, Klaviyo, Ryze AI, Airtable, Cal.com and Calendly. See integrations. |
Boundaries
- Inside your workspace only. An agent reads and writes through the same workspace isolation and private-task rules as everyone else. It cannot see another workspace's data.
- Tools are granted, not assumed. An agent can only reach an integration you have connected and ticked for that agent.
- Metered by actual usage. Runs spend workspace credits at what the work really cost — model usage, web searches and scrapes — never a flat fee per job. Billing breaks spend down per agent, so you can see which one is expensive.
- Fail-closed. Credits are checked before a job starts and before each step. An empty wallet stops the run and says so in a comment rather than running up a bill.
- Everything is on the record. Work happens as task comments and job steps, so there is always a readable trail of what an agent did and why.
Agents vs the AI assistant
They are different tools. The Kyo AI assistant in the Library answers questions from your knowledge base — it reads, it doesn't act. An agent does the work: it changes records, posts messages, and produces something you can review. Both are metered against the same credit wallet.
Through the API
Agents and their instructions are readable through the
REST API at /v1/agents under the
agents:read scope, and through the
MCP server as kyo_list_agents and
kyo_get_agent. Agent rows are written by Kyo itself, so the resource is read
only — create and edit agents in Settings → Agents.