With this, your mind can take part in your Slack threads like a colleague who knows when not to talk.

Why. We live in Slack. An agent that only answers when you @-mention it is a search box. One that reads the thread can notice that you are stuck.

How it works. Each thread gets an agent of its own, at /agents/slack/<team>/<channel>/<thread>. The agent has no special Slack toolkit. It calls Slack's Web API itself, through one method of the mind's code, so anything Slack can do, the agent can do:

await itx.config.slack("chat.postMessage", { channel, thread_ts, text });

Our files, to copy the shape of:

  • slack.ts: who may give an agent orders (a full member of the workspace whose confirmed email is a member of the mind; never a bot, a guest or another workspace), which thread agent a message goes to, and slackForAgent, which marks every non-member's words as a quote when the agent reads the thread back.
  • slack.md: when to talk. Chime in when asked or when you have something concrete. Do not repeat people. Otherwise stay silent. Show Slack's "is thinking" status while working.
  • worker.ts: the slack/webhook-received case, which hands each message to its thread's agent, and the slack method that agents call.

What it needs. A Slack workspace, linked by a member through iterate Cloud's Slack app; the token lands in /secrets/slack and never leaves the vault. Every member message in a thread is one agent turn, so it costs a model call per message.

What surprised us. "Stay silent" is the hardest instruction we gave it, and the most important. An agent that answers everything is one you mute by Friday.