Send to agent
Approving work and acting on it are separate steps. Send to agent connects them: from an approved Solution plan or a decided Decision, open the Compass in-app agent in a new conversation that already knows what was approved, instead of describing it again from memory.
This is an entry point to the existing in-app agent, not a separate assistant. The agent still runs with your own workspace permissions and the workspace's enabled capability packs.
Where it appears
Send to agent is deliberately limited to work that has actually been approved:
- Solution plans. In the Solution detail panel's Plan & Discussion section, next to Approve and Reject, once the current plan is approved. A pending or rejected plan does not show the action.
- Decisions. On a decided Decision, in the recorded-decision banner, when the chosen outcome was an approval. Outcomes of Request changes or Reject do not show it.
Release authorization and legacy system decision records share the same decided banner but do not offer this action. It applies to ordinary tracked Decisions only.
What happens when you use it
Selecting Send to agent opens a new agent conversation with two things prepared:
- A context chip above the message box, naming the plan or decision and linking back to it.
- A suggested instruction already typed into the message box.
Nothing is sent automatically. The suggested instruction is ordinary editable text — replace it entirely if what you want is different from what Compass guessed. An agent turn only starts when you send it, so arriving here never consumes turns or counts against usage limits on its own.
Once you do send it, the turn runs on the server and you can leave — see Background agent runs.
Dismiss the chip with its × to send the message without the attached context. Dismissing the chip does not clear anything you have typed.
What the agent receives
The attached context is resolved fresh when you send, from the plan or decision itself — not from the link you followed. A long decision context is shortened, with a pointer back to the source for the full text.
Context is attached to the first message only. Later messages in the same conversation do not re-send it, so a long conversation does not keep paying for the same attachment. To attach it again, or to attach a different plan or decision, start a new conversation from that item.
Your conversation history records only the message you actually typed. The attached context is never stored as though you had written it.
If the work changes before you send
A plan can be rejected, or a decision reopened, between opening the agent and sending your message. When that happens the message still sends — without the attached context, and without an error. Compass does not attach context from work that is no longer approved.
The same applies to a stale or shared link: following one for an item you cannot access, or that no longer exists, opens an ordinary empty agent conversation with no chip and no prefilled text. Access is always evaluated against your own workspace membership, so a link does not grant anyone visibility they did not already have.
Reopening an existing conversation never prefills the message box or restores a chip, even if that conversation originally started from an approval.
Viewing Compass inside Geode
If Compass is open inside Geode's Web Viewer, Send to agent becomes a small menu instead of a single button: Built-in cloud agent (everything above, unchanged) and Geode, which hands the same context to a local Geode Agent Threads runtime instead of Compass's own in-app agent. Outside Geode — i.e. every normal browser — nothing changes; the single-button behavior above is exactly what you see.