Delivery

An agent that produces a perfect brief nobody reads has failed. Delivery is a first-class part of the spec, not an afterthought.

Slack

The most common destination. The Delivery card lists the channels an agent posts to, as #channel-name. An agent can post to more than one.

Guidance that survives contact with real teams:

• One channel per agent, at first. A shared #taizen channel becomes unreadable within a week.
• Start private. A channel with the two or three people who care lets you fix tone before an audience forms.
• Match the channel to the action. If the output needs a rep to do something, it belongs where that rep already works, not in an ops channel.

Requires the Slack integration.

Email

The Delivery card lists individual email recipients. Email suits:

• daily and weekly digests people read outside Slack
• output for managers who live in their inbox
• anything long enough that a Slack message would be a wall of text

CRM write-back

Agents can write to your CRM: deal fields, next steps, close dates with the reason captured from the call, activity logging. This is the Auto CRM Update pattern, and it's usually the fastest visible win because it removes work reps currently do by hand.

Two cautions:

• Write-back needs write scopes. A read-only key means the run reports that it couldn't write rather than pretending it did.
• Watch it first. CRM fields are the record other systems read. Run it on demand and read the output for a week before you leave it unsupervised — see validation and review.

Live Apps

An agent can publish an HTML page at a stable URL, which keeps its data live between runs — a pipeline dashboard, a sales activity report, a deck for a customer. Later runs update the page rather than producing a new artifact each time.

See Apps.

Files

Runs can produce files — PDF, DOCX, PPTX, XLSX, CSV, HTML, images. They're attached to the run and visible on the run in the Timeline, so a reviewer can open what was generated.

"Framework default"

If the Delivery card says Framework default, delivery hasn't been configured and the agent uses the default for its type. That's fine for testing. Before you rely on an agent, set the destination explicitly — a default destination is how output ends up somewhere nobody is looking.

When delivery fails but the run didn't

There's a distinct status for exactly this: the agent did its work, and posting the result failed (a revoked Slack token, a channel it was removed from, an expired CRM token). The output isn't lost — it's on the run. Reconnect the integration and re-run.

See runs and debugging.