The four-bucket model

For each item the agent encounters, exactly one of these:

Bucket	When	What the user sees
1. Auto-archive	Obvious noise (newsletters, mass mailers, status alerts), items already handled by someone else, items needing no action	One-line FYI in the batched summary, always with a reason
2. Auto-handle	The agent can fully complete the work itself — mark a CRM task done, schedule a follow-up reminder, archive after the other side already replied	One-line FYI in the batched summary
3. Draft → approve	A reply is needed	Reviewed one by one in the batched output
4. Park	Ambiguous, or the user's judgment is required	Left in inbox with a one-line reason

Steps

Sweep all connected surfaces. Pull unread / unactioned items from email, Slack DMs, LinkedIn DMs, the Swan inbox, and the CRM task queue. Count them. If above ~30, switch to swan-execute-code — load the inbox snapshot into a frame and process in passes.
Classify silently into the four buckets. For each item:
- Read the message.
- Investigate first when facts are needed — check the CRM for prior touches and the account record, check the subscription, search product documentation, pull recent Slack threads where the customer was discussed. If the question is about a bug, product behavior, or code-level issue, route to the org's internal investigation/auditor agent when one is available (these typically have access to logs and the codebase; that context lands in the draft).
- Decide the bucket. When uncertain → park, don't auto-archive.