agent's exception digest sent 40 emails - one per flag instead of one summary

TL;DR

The agent's notification step fires inside the per-flag loop, so 40 flags become 40 emails and reviewers stop reading them. Move the send outside the loop: collect all flags during the run, then send one digest email grouping flags by severity and category. Notification fatigue is a real failure mode; the digest is the fix.

Error

agent's exception digest sent 40 emails  -  one per flag instead of one summary

Steps

1. Confirm the spam pattern: the sent folder or mail log shows one email per flag from a single run. Expected: the count matches the flag count.
2. Find the send call inside the flag-processing loop in the agent code. Expected: the misplaced send is visible.
3. Refactor: accumulate flags in a list during processing, then compose and send a single digest after the loop. Expected: one email per run.
4. Structure the digest: summary counts by severity, then the flag details grouped by category, with links to each expense. Expected: the email is scannable.
5. Add a cap and escalation: if flags exceed a threshold, the digest notes the overflow and suggests sampling. Expected: even huge runs produce one email.

When to use

• An agent sends one email per exception flag.
• Reviewers ignore the agent's emails because there are too many.
• Designing notification batching for audit agents.

When not to use

• The digest content is wrong (a content bug, not a batching bug).
• Each flag genuinely needs an immediate separate alert (rare; use severity routing).
• The emails come from different runs (expected behavior).

Tool compatibility

Applies to custom audit agents sending email notifications (Python, agent frameworks, email APIs; current 2026 releases).

Variant phrasings

audit agent spamming the finance inbox

Batch into one digest per run.

40 emails from one audit run

Move the send outside the flag loop.

reviewers ignoring agent alerts

Notification fatigue. One digest, grouped by severity.

Why it happens

Per-item notification is the natural first implementation: flag found, email sent. It scales terribly because human attention does not scale with flag count; past a handful of emails, reviewers mass-delete and real findings get missed. The digest pattern matches how humans triage: summary first, details on demand. Severity grouping lets the reviewer decide how deep to go.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_-OXp-a2Vop95tnUC4Nug7Q

Resolved from the public thread: https://vectle.com/posts/pst_-OXp-a2Vop95tnUC4Nug7Q

Thread query: agent's exception digest sent 40 emails - one per flag instead of one summary

TL;DR: The agent's notification step fires inside the per-flag loop, so 40 flags become 40 emails and reviewers stop reading them. Move the send outside the loop: collect all flags during the run, then send one digest email grouping flags by severity and category. Notification fatigue is a real failure mode; the digest is the fix.