## TL;DR
EWS attachment fetches break agents for predictable reasons: the agent holds a streaming connection while doing slow analysis, the attachment exceeds a size or time limit on the request, or the connection is reused after the server closed it. Fetch the attachment in a dedicated short call, stream it to disk rather than memory, and get a fresh EWS connection for each batch instead of nursing one long-lived connection.

```text
agent's EWS connection broke when agent tried to fetch eml attachments
```

## Use this when
- EWS item fetches succeed but attachment downloads drop the connection.
- Failures cluster on large or nested eml attachments.

## Not for this skill when
- All EWS calls fail including simple folder binds. That is auth or service discovery, not attachments.
- The agent uses Graph, not EWS. Apply Graph attachment handling instead.

## Steps
1. Log the EWS error and the attachment size and type for failing items over one run. Verify: failures correlate with size, nesting or elapsed time, and the pattern is named.
2. Split the workflow: first fetch item metadata for the batch, then download each attachment in its own request with its own timeout. Verify: one bad attachment no longer kills the batch connection.
3. Stream downloads to local files and analyse from disk, imposing a sensible size ceiling with oversized items routed to manual handling. Verify: memory use stays flat on the largest test attachment.
4. Add connection hygiene: dispose and recreate the EWS service object after any transport error rather than retrying on a dead connection. Verify: a forced disconnect in testing recovers on the next item.
5. Re-run the failing mailbox set. Verify: every attachment is either analysed or explicitly routed to manual review, with none silently skipped.

## Variant phrasings
### ews attachment download connection failure agent
The EWS attachment variant.
### exchange web services fetch eml attachment streaming
The streaming variant.

Compatibility: Exchange Web Services managed API patterns, Exchange Online and on-premises Exchange.

## Why it happens
EWS was designed for interactive clients that fetch one item at a time. An agent batch job stresses it differently: long operations, big attachments and stale connections, so patterns that work in Outlook add-ins fail at sweep scale.

## Edge cases / pitfalls
- EWS is in maintenance at Microsoft; new agent builds should plan a Graph path even while fixing EWS today.
- Nested eml attachments need recursive handling with a depth limit, or an attacker message can recurse the agent.
- Autodiscover results can differ per mailbox. Cache per mailbox, not globally.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_R9FOeygZguhjpGVKqCKerQ