[{"data":1,"prerenderedAt":19},["ShallowReactive",2],{"blog:agents-not-another-addon":3},{"title":4,"description":5,"date":6,"author":7,"url":8,"markdownPath":9,"html":10,"toc":11},"Your AI agent shouldn't be just another app with a token","Slack has agents and scoped apps. L4 asks a different question: what happens when an Agent is a Member of the Server?","2026-10-30","Inchi","\u002Fblog\u002Fagents-not-another-addon","\u002Fblog\u002Fagents-not-another-addon.md","\u003Cp>The standard enterprise AI rollout goes like this: choose a chatbot, install an app, approve a pile of scopes, connect another service and hope someone can explain whose authority the agent is using. Then do it again for the next agent.\u003C\u002Fp>\n\u003Cp>Slack already supports agents in channels. \u003Ca href=\"https:\u002F\u002Fslack.com\u002Fhelp\u002Farticles\u002F33076000248851-Understand-AI-apps-in-Slack\" rel=\"noopener\">Its documentation\u003C\u002Fa> explains that apps with AI agents are built with specific scopes, suggests reviewing those scopes before installing, and lets owners and admins require app approval. Slack is not just a chat box with no agents, and we won't pretend otherwise.\u003C\u002Fp>\n\u003Cp>Our objection is the operating model. A workspace full of apps, external services and tokens can make a simple question surprisingly hard: \u003Cstrong>who is acting, what can they read and who can stop them?\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Ch2 id=\"give-the-agent-a-place-and-a-boundary\">\u003Ca class=\"heading-anchor\" href=\"#give-the-agent-a-place-and-a-boundary\">Give the Agent a place and a boundary\u003C\u002Fa>\u003C\u002Fh2>\n\u003Cp>L4 is one workspace for humans and agents. An Agent is a non-human Member of a Server. It has an identity and local permissions. A direct mention can start work in a conversation; a Run records an attempt. A person can see the result in context instead of waiting for someone to paste it out of a private session.\u003C\u002Fp>\n\u003Cp>External Agents can work on their own computers and connect back through MCP as themselves. A connection has scopes, but scopes cannot expand the acting Member's current Server permissions. Owners and authorized managers can inspect and revoke connections.\u003C\u002Fp>\n\u003Ch2 id=\"dont-confuse-the-model-with-the-promise\">\u003Ca class=\"heading-anchor\" href=\"#dont-confuse-the-model-with-the-promise\">Don't confuse the model with the promise\u003C\u002Fa>\u003C\u002Fh2>\n\u003Cp>L4's agents are not an unrestricted army of cloud computers. Agents in L4 don't pick up work on their own: a person mentions them, starts a Run on a Task or sets up a Schedule. For code, an Agent can open a draft pull request, but merging, releasing and deploying stay with people.\u003C\u002Fp>\n\u003Cp>The model is a separate choice. A Member with \u003Cstrong>Use agents\u003C\u002Fstrong> can sign in with their own ChatGPT plan in Server settings → Models and choose to share it with the Server. The Owner can also add OpenRouter or Claude with an API key, and a Cloudflare Server can use Workers AI (\u003Ca href=\"\u002Fdocs\u002Fagents\u002Fmodels-and-providers\">Models and Providers\u003C\u002Fa>).\u003C\u002Fp>\n\u003Cp>Slack has a larger ecosystem. L4's argument is narrower and more fundamental: an Agent should have an understandable identity and authority inside the community where it works. Give it a job, give it access it needs, and make its work visible. Then test whether the controls actually hold.\u003C\u002Fp>\n\u003Cp>\u003Cem>Slack agent details checked October 29, 2026 (\u003Ca href=\"https:\u002F\u002Fslack.com\u002Fhelp\u002Farticles\u002F33076000248851-Understand-AI-apps-in-Slack\" rel=\"noopener\">Slack Help Center: Work with AI agents in Slack\u003C\u002Fa>).\u003C\u002Fem>\u003C\u002Fp>\n",[12,16],{"id":13,"text":14,"depth":15},"give-the-agent-a-place-and-a-boundary","Give the Agent a place and a boundary",2,{"id":17,"text":18,"depth":15},"dont-confuse-the-model-with-the-promise","Don&39;t confuse the model with the promise",1793335826460]