Agent API returns empty responses (HTTP 200) on generation calls

We have an app deployed in abp-sandbox for “offshore regulatory assistance” (Flows + KG-RAG agent) on the Agent API (`az-ams-sp-002`, `cdf-version: beta`). While developing document export, I've encountered a reproducible failure pattern:

Generation calls with complex instructions and/or long input intermittently return HTTP 200 with empty `content.text` — no error signal. The failure threshold shifted within hours on identical configuration: a 28.7K prompt succeeded at 18:15, 25K failed 12/12 an hour later, 15K succeeded at 21:49. Compression/summarization calls of similar size succeeded in the same window where generation calls failed.

A report with the full test matrix is attached.

Questions for Cognite (prepared with AI assistance):

1. Could the endpoint return an explicit error/finish-reason instead of an empty 200?
2. Are there documented practical limits for input size × instruction complexity?
3. What is the roadmap status for streaming or async chat?
4. Does a plain LLM endpoint without agent/RAG exist (or is one planned) for tasks that don't require retrieval?

15.07.2026
The empty-response issue is now deterministically reproducible — 10/10 empty with a 28K prompt vs. 33/33 OK with a short prompt, all within 20:01–21:32 UTC on 14 July (call ids and timestamps in the report). The report also documents a separate cognite-sdk hang triggered by these responses; please route that part to the SDK team.