Working with the Agent in Slack

When the agent replies, how threads work, and the task and fork commands

Talking to the agent feels like talking to a teammate in the channel. The one thing worth learning is how you address it, because that decides what context the agent brings to your request.

When the agent replies

The agent sees messages in every channel the app has been invited to. It replies:

• When you @mention it.
• When you reply in a thread the agent started or was addressed in. No new mention is needed.
• When it judges a reply useful for any other channel message. The agent batches these ambient messages for a couple of seconds, then decides whether to answer or stay quiet.

Three ways to start a conversation

At the top level of the channel, the word right after the mention picks the context mode.

You write	What happens	Use it for
@Agent <request>	Continues the channel’s main session, the shared running context for the channel.	Everyday questions that build on what the channel has discussed.
@Agent task <request>	Starts a fresh session in a new thread, with no prior conversation context.	Independent work that the channel’s history would only clutter.
@Agent fork <request>	Starts a new thread session seeded with a snapshot of the channel’s context so far. Later channel messages stay out of the fork.	A side investigation that needs what the channel already knows.

@Agent what did we decide about the retry policy?
@Agent task draft a migration plan for the billing webhooks
@Agent fork dig into why yesterday's import failed

The command word must come right after the mention, at the start of a top-level message. A colon after the mention works too, as in @Agent: fork ….

Reach for task when the request stands alone, and fork when it depends on the conversation so far. Both keep the channel’s main session focused.

Threads

Each thread has its own session, so work in one thread stays separate from the channel and from other threads.

• Agent threads. Replies inside a thread the agent owns continue that thread’s session.
• Implicit forks. Replying in a thread under one of the agent’s channel-level messages forks the channel’s context into that thread.
• Thread context wins. Inside a thread, task and fork are ordinary words. The agent continues the thread’s session.

Sessions over time

Every session, whether the channel’s main session or a thread session, keeps its own context and its own sandbox.

• Idle sessions sleep and wake when someone writes to them again.
• Long conversations are compacted. When a session approaches its compaction threshold, the agent summarizes older conversation to make room. The default threshold is 64,000 tokens. See Models and cost.
• Closing a session discards its resumable context and its sandbox files. Slack history and anything the agent saved to Moxn remain. You close thread sessions from the agent page. See Operations.

Because sessions are temporary, ask the agent to record anything worth keeping in Moxn. With the default memory instruction on, it records decisions and findings on its own. See Knowledge and memory.

Agents work in public channels and in private channels the app is invited to. Shared channels (Slack Connect), direct messages, and group DMs are not supported yet.