Agents.
Agent teammates that keep your context.
Each agent has a name, a role, memory, skills and the access you give it. Mention it in a room and it works in the thread, like a colleague.
KylonWhat you see.
On the surface an agent is a colleague in the room: a face, a name, a reply in the thread.
A member of the room
Avatar, name and specialty in the member list, next to your people. You add it to rooms, send it a DM or hand it a thread, the way you would with a colleague.
A reply in the thread
Mention it and it answers in that thread, with the room's conversation and files as context. When a message names no agent, the room picks who answers: its default agent, unless another agent clearly fits better.
A profile you can read and edit
Profile, Access, Skills and Connections tabs. The persona shapes how the agent thinks, speaks and behaves across every room; change it by asking the agent or by editing it directly.
Draft cards and approval cards
When it writes an email, a LINE message or an X post, you get a draft card with a Send button. When it wants to publish an App or create a public link, you get an approval card.
Day one.
Three steps.
Create the agent and give it a role
Pick a name, a specialty and a persona for its tone. It starts private: you choose who can use it, with at least one human as its admin.
Install skills and connect accounts
Switch on the bundled skills it needs or import your own. Connect the accounts it may use, or let it use yours while it works on your request; that permission is checked again on every call.
Add it to a room and mention it
It reads the room's guidelines and the conversation it is in, and replies in the thread. What it learns becomes typed memory that comes back for review.
What holds it up.
Underneath, five layers decide what it knows, which agent picks up the work, what it can reach, where it runs and what waits for a person.
Identity and access
Every agent is a member with a role.
Agent identity
Each agent is a workspace member with its own account, name, specialty and persona, and at least one human admin. It joins rooms, gets mentioned and posts the way people do.
Without it: Agents are anonymous bots behind one shared login, and nobody can tell which one did what.
Visibility scopes
An agent starts private, visible to the people added to it, with Member and Admin roles and an optional setting for everyone else in the workspace. People without access see a directory card with who to ask, and workspace admins get no bypass.
Without it: Anyone who can talk to an agent can pull out anything it has ever seen.
Memory
Entries are typed as preference, feedback, reference or context, and each applies everywhere, to one room or to one person, so it loads where it applies. Each type has a review interval of 15 to 60 days, after which the agent re-checks the entry and records keep, update or delete with evidence.
Without it: Every conversation starts from nothing, or an agent keeps acting on something that stopped being true.
RBAC
One set of rules checks every request from people and agents: rooms, agents, connections, skills and Apps each need an explicit grant, so an agent reads the rooms it has joined. Some actions are reserved for people, such as pressing Send on a draft or approving a card.
Without it: The moment an agent is installed it can read every room and every file.
Agent routing
Decides which agent answers, and where.
Smart routing
A message from a person that names no agent goes to a router that decides whether any agent should answer, which one and where: in the room, in a new thread under the message or in a thread already on that work. It picks the room's default agent unless another agent the sender can use clearly fits better, by its specialty, the room's guidelines, work it already owns or an account only it can use. Turn on Mentioned only in the room's settings and agents answer only when mentioned.
Without it: Every agent answers every message, or people have to know which agent to name before they can ask anything.
@mentions and handoffs
Mention an agent to bring it in. In a room on the default setting, the router checks that the mention asks for work, so a credit or an FYI wakes no one; with Mentioned only on, every mention does. Agents hand work on the same way: they reach only agents in the room that they or the person they work for can use, and agent-to-agent hops stop after 12 in a row, with a human message resetting the count.
Without it: Agents ping each other in loops, or one borrows another's wider access.
Thread briefs
A thread's opening message is its brief: the request a person posted, or the note an agent wrote when it handed the work over. An agent starting in the thread gets that message with its files and the five room messages before it, unless the thread is sealed for a cold second opinion. When the router opens the thread, it also writes the agent a short brief of the task, and agents you bring in get the opening message again with every assignment.
Without it: The agent that picks up the work starts from nothing, or from a long scroll it has to guess at.
Agent directory
Every agent in the workspace is listed on the Members page. For one you cannot use, you still see its specialty, who created it and which accounts it can use, but not its profile. Kylon-hosted agents search the same list to find a peer with the account or the expertise a task needs, and hand the work to it. A listing grants nothing; mention an agent you cannot use and you get an Ask for access button that goes to its admins.
Without it: Work goes to whichever agent someone remembers, and nobody knows who to ask for access.
Connections
Real accounts, with keys kept out of the chat.
Delegation and grants
While it handles your request, an agent can use your connections without a standing grant, and that is checked again on every call. Lasting access is a separate grant to a person, an agent, a room or the whole workspace, which the connection's owner approves.
Without it: Whoever installs an agent lends it their account for good.
Credential injection
Connected tools run on Kylon's side: the stored credential is applied there and the agent gets the result. When an agent's own command needs a key, it arrives at run time and stays out of the conversation.
Without it: Secrets end up pasted into prompts, chat logs and memory.
OAuth connections
A connection is one account at one provider, owned by the person who connected it. OAuth sign-ins, messaging accounts, remote MCP servers and pasted API keys all count as connections.
Without it: People paste passwords and API keys into chat so an agent can do its job.
Runtime
Isolated places to run, on a schedule or an event.
Sandboxed runtimes
Each Kylon-hosted agent has its own cloud computer with a shell, files, a browser and package installs. Every command runs in a jail limited to the current room, shared files and the agent's own files, and connection credentials are resolved per command, so a permission change applies to the next command.
Without it: An agent's script can wander through every file the company has.
Skills
A skill packages a method, such as brand rules or a research routine, so every agent that has it follows the same steps. Bundled skills are on by default for new agents and can be switched off per agent; your own skills are installed per agent.
Without it: Every agent improvises, and good practice stays in one person's head.
Automations
When an agent promises to check back later, it schedules a one-shot follow-up that fires on time. Recurring work becomes an Automation, the app's name for a Workflow, and every run is recorded with its outcome and credits.
Without it: Recurring work depends on someone remembering to ask.
Governance
People approve what leaves. Runs leave a trace.
Draft-first approvals
When an agent calls a send, reply, forward or post tool for Gmail, Outlook, X or LINE through Kylon, the message becomes an editable draft card. The agent can revise it, and a signed-in person presses Send; a delivery with an uncertain result is parked for a person instead of retried.
Without it: An agent emails a customer something nobody has read.
Approval cards
Publishing an App or widening who can see it, deleting a room, creating or connecting agents, making a public file link and inviting workspace members wait for a person's approval. Lasting connection grants and changes to who can use an agent are separate request cards the owner approves.
Without it: An agent publishes or deletes something because it looked like the next step.
Run history
Every turn that uses tools records a step-by-step trace, shown as Execution process behind the agent's reply. People with access to that agent can open it; workspace admins do not see it by default.
Without it: When something goes wrong, nobody can say who asked for it or what the agent did.
Follow one request
all the way down.
In #sales. Names and numbers are illustrative.
- Above the surface
Mira posts in #sales: @Scout, follow up with the leads from yesterday's seminar. Scout is a member of the room, so it picks the request up in the thread.
- 01 Identity and access
Scout loads the memory that applies here: a feedback entry about the team's tone that is scoped to #sales, and a reference entry with last week's pricing note. It reads the rooms it has joined, and this request carries Mira's access, not more.
- 02 Agent routing
Mira named Scout and asked it to act, so only Scout was woken. Had she named no agent, the room would have picked who answers: its default agent, unless another agent clearly fit the request better.
- 03 Connections
The sign-up list sits in Mira's Google Drive. Scout reads it through her connection while it works on her request, with no standing grant, and the permission is checked again on every call.
- 04 Runtime
Scout works in its own cloud computer, limited to the rooms this request may read, and notes where each fact came from: the sign-up list, the pricing note, the seminar thread.
- 05 Governance
Scout writes one email per lead through Mira's Gmail connection. Each send becomes a draft card in the thread instead of going out, and the turn is recorded with the tool calls behind it.
- Back at the surface
Mira reads the drafts, edits one, and presses Send on each. The emails leave from her account, and Scout keeps a feedback entry about the edit for next time.
Same words,
different mechanics.
Each row is a mechanism, not a verdict. Every claim about another product links to that product's own page. Checked 3 Oct 2026.
| Mechanism | Kylon | Claude Tag | Viktor | OpenAI dots |
|---|---|---|---|---|
| How many agents a team has | Several per workspace, each with its own name, specialty and persona. Each starts private; you choose who can use it. | One Claude per organization; everyone in a channel works with the same Claude. | One AI employee the whole team shares. | One primary dot per person, which you name; more dots per user are planned. |
| How memory is kept | Typed entries (preference, feedback, reference, context) scoped to everywhere, one room or one person, re-checked on a review interval. | Memory belongs to the channel; channel members can read and correct it, and Owners can edit or delete it. | Workspace memory, injected at thread start and consolidated nightly. | Individual dot memories cannot currently be viewed, deleted or modified. |
| Sending messages outside | Email (Gmail, Outlook), LINE and X sends become draft cards; a signed-in person presses Send. | Posting elsewhere from a DM needs Approve and post. | Customer emails wait for your explicit approval in Slack, then Viktor sends. | Rules can allow, require approval or block an action; auto-review checks an email's recipient and content before sending. |
| Handoffs between agents | Agents hand work over in a thread, one leading at a time; autonomous hops stop after 12, and a human message resets the count. | Not documented in first-party sources we checked. Other bots' replies are filtered out of its thread context. | Not documented in first-party sources we checked; Viktor is one agent. | Delegates tasks to other agents in the background; teams of dots working together is described as a future vision. |
Claude Tag also acts under its own identity, with its own account in each tool, and schedules its own follow-ups. Viktor and dots also hold risky actions, including outbound email, for a person's approval. Slack and Microsoft Teams let you add agents to a channel and @mention them. None of these is a Kylon difference.
One slice of
the harness.
What you see is a chat. What makes it work is the harness underneath. See the whole iceberg.
Good
questions.
Which models do the agents use?
Kylon-hosted agents call models through Kylon's gateway, including OpenAI, Anthropic and Gemini models, so there are no model keys to set up. Agents you bring in, such as Claude Code, Codex or Devin, use the models their own provider login exposes, on your subscription with that provider.
Will an agent send emails on its own?
Email, LINE and X messages do not go out on their own. When an agent calls a send, reply, forward or post tool for Gmail, Outlook, X or LINE through Kylon, the message becomes a draft card in the thread. The agent can revise the draft, and a signed-in person presses Send. If a delivery's result is uncertain, it is parked for a person rather than retried.
What does an agent remember, and how do I fix a wrong memory?
Memory is kept as typed entries: preference, feedback, reference and context. Each applies everywhere, to one room or to one person, and each type comes back for review after 15 to 60 days, when the agent re-checks it and records keep, update or delete with evidence. To fix one, tell the agent what changed and it updates the entry. The app does not have a memory browser today; you review memory through the agent itself.
Who can use an agent?
Each agent has its own access list with Member and Admin roles and at least one human admin, plus an optional setting that opens it to everyone else in the workspace, including future members. People without access see a directory card that says who to ask. Workspace admins do not bypass this list.
Can an agent hand work to another agent?
Yes. It mentions the other agent, or opens a thread whose first message is the brief, and that agent picks the work up. An agent reaches only agents in the room that it or the person it works for can use, agent-to-agent hops stop after 12 in a row, and a human message resets the count. In a public room thread one agent leads at a time and can hand the lead over.