The choice between a Slack agent, Slackbot and Agentforce comes down to who owns the machine the work runs on. Slackbot and Agentforce answer from what Slack and Salesforce let them read. Kortix is the open-source AI Operating System, and its Slack agent runs a session on a cloud computer with your connectors and your approval rules.
What each one runs on, what it can reach, and where its configuration lives.
kortix connectors policy set --default risk --apply # or allow_all; without --apply it edits the local kortix.yamlkortix connectors policy add "gmail.send_*" ask --condition to='*@example.com' # project-wide rule; k!=v negateskortix connectors policy gmail set send_email block # connector rule: allow, ask or blockkortix connectors sensitive gmail on # or off
Source: https://kortix.com/docs/connect/connectors
Slackbot has become a personal AI agent. It ships inside Slack with no setup, it reads your messages and files, and it sees only what you can already see. Ask it to summarize a long thread, draft an update in your voice, or pull the numbers out of a 40-page deck, and it answers in place.
That covers reading and writing inside the product. Slack's announcement describes Slackbot doing catch-up, drafts and file analysis. It does not describe an agent that runs a job on a machine of yours, calls your CRM, or stops before a write to ask a person. When you want that, you are shopping for a runtime, and Slack becomes the surface you talk through.
Agentforce is Salesforce's product, and Slack documents the split. You build an agent in Agentforce Builder. Then you connect a Salesforce organization to Slack, map accounts, and have an admin review the permissions the agent will hold before installing it. A person without a Salesforce account needs a provisional Salesforce license to use one.
The grounding is the point. An agent reads Slack conversations and Salesforce CRM records together, so a support lead gets a customer brief in the channel where the escalation is happening. A team already running Salesforce gains reach that Slackbot cannot give it. A team that is not on Salesforce now has two platforms to license. Its agent's configuration also lives in Salesforce, not in a repo the engineering team reviews.
Dust is a multiplayer AI platform: people and agents share one knowledge base, and you call an agent from Slack the same way you call it from Dust's own app. Agents pull from Notion, GitHub, Google Drive, Salesforce and Zendesk, so questions that need material outside Slack get a real answer.
You build and configure the agent in Dust. For a company whose knowledge sits across five tools, and whose team wants one governed place to ask, that model fits. The agent answers from Dust's platform, and the knowledge base it reads lives there too.
Three things decide whether the agent can finish a job. The first is a computer: a session that runs somewhere other than the thread, with a working directory, files and a terminal. In Kortix a Slack thread is one session, and a mention in a channel reaches it. The agent replies in the thread with live progress while it works on a cloud computer.
The second is governed reach. A connector gives a project access to one app, MCP server or API, and Kortix holds the account, applies the project's policies and records every call. A call that needs approval is held: it returns 202 with status: "pending_approval" and the parameters the provider will receive, and nothing ran. The card posts in the Slack thread. A person approves or denies that exact request, and a changed argument needs a new approval.
The third is a way back. Session work reaches main through a change request. Merge is default-deny for agents, so what ran overnight is reviewable in the morning. Approval gates you set. Off until you set them, which means a new project runs open until someone turns on Ask or Block for the actions that matter.
None of this needs a Salesforce license or a second system of record. Connectors live in kortix.yaml, agents live in the repo, and you run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.
Buy the smaller thing when the job is smaller. A team that wants summaries, recaps and drafts inside Slack, with nothing to install or govern, should use Slackbot. It ships inside Slack and already holds the context. A Salesforce shop that wants CRM-grounded answers in the account channel should use Agentforce. The license it already pays for is the hard part. A company that wants one shared knowledge layer over Notion, Drive, GitHub and Zendesk, with agents called from Slack, should look at Dust.
Kortix earns its place when the agent has to act. Send the email, open the ticket, update the record, then land the change in a repo someone reviews. That work needs a computer, a governed credential and a record of each call. A chat surface provides none of the three.
main through a change request.Connect Slack to a project, invite the bot, and mention it with a task.
Approval gates you set · Self-host is free · 3,000+ apps in a click