Video Agent (v3) sessions failing at progress 0 with no error code — intermittent, worse with longer prompts

[satkar-adh]

Description

Subject: Video Agent (v3) sessions failing at progress 0 with no error code — intermittent, worse with longer prompts

Account API key ending: ZVMn
Date of tests: 2026-09-01, ~01:00–04:00 UTC

Summary

POST /v3/video-agents is accepted (HTTP 200, session_id returned), the agent
posts an opening message, then the session goes to status "failed" with
progress 0 — before any rendering starts. No reason is returned anywhere.

This is intermittent and correlates with prompt length: short prompts almost
always succeed, prompts of ~2.5k characters and up frequently fail. The SAME
prompt succeeded at 01:36 and failed at 03:00 with no changes.

No error information is returned

GET /v3/videos/{video_id}
{"data":{"created_at":...,"id":"...","status":"failed","title":""}}
-> no failure_code, no failure_message, despite
https://developers.heygen.com/reference/get-video documenting them

GET /v1/video_status.get?video_id=...
{"code":100,"message":"Success","data":{... "status":"failed","error":null ...}}

GET /v3/video-agents/{session_id}
keys: created_at, messages, progress, session_id, status, title, video_id
-> no error / failure_code / failure_message field at all
-> last message is the agent's own: "I'll start by creating a brand kit ...
then I'll draft the full blueprint"

Failing session IDs

4534480298564a76b31b1500534719db 290-char prompt, portrait -> failed t+60s
591999b1c13c4dd6988bbcaa72d5280d 3592-char prompt, portrait -> failed
bcc376b0544d44319c2baa67d336d5df 3592-char prompt -> failed
dd484f35feee4fda9afdbafd9bc3b180 5745-char prompt -> failed
8512749f257a4035b28ccc2095dc811b 5745-char prompt -> failed

Succeeding sessions, same key, same hours (for contrast)

a503824addd64bc7a219ea058e121f07 105-char prompt, portrait -> completed
88e940dda8ab4105a8069184fd980f90 105-char prompt, landscape -> completed
ab48ebe7c40047c9b60fe61879191c1e 3592-char prompt -> completed
a391ee1ff48c43bc8f6085d19f69f620 2394-char prompt -> completed 01:36
(the identical 2394-char prompt FAILED when resubmitted at ~03:00)

What we ruled out on our side

• Quota: 342–465 remaining throughout; exhaustion would return 429 at submit,
  and these are accepted with a session_id.
• Orientation: identical prompt submitted as portrait and landscape — both
  completed. Not orientation-specific.
• Prompt content: a prompt containing clinical/injection wording failed, but the
  same content at 2,394 chars completed, and a completely unrelated prompt of
  similar length also failed.
• Request shape: mode=generate, orientation, voice_id (VpwKdM1cvA7NWyJVaA1Y —
  confirmed present in /v2/voices). With and without callback_url: no difference.
• Prompt structure: splitting one long scene into several made no difference.

Questions

1. Is there a documented or undocumented prompt-size limit for /v3/video-agents?
   If so, what is it, and can the API reject oversized prompts at submit time
   (4xx) rather than accepting them and failing silently later?
2. Why are failure_code / failure_message absent on GET /v3/videos/{video_id}
   when the reference documents them? Is there another endpoint carrying the
   reason?
3. Were there known Video Agent incidents in this window? status.heygen.com
   showed all-operational throughout.

Steps to reproduce

Steps to reproduce

1. Send a video generation request to the Video Agent with a short prompt —
   two or three sentences, portrait orientation, one of your standard voices.
   It is accepted, and a few minutes later the video finishes normally.

2. Send the same kind of request again, but with a longer prompt — roughly
   3,000 to 6,000 characters, the sort of length a detailed brief produces.
   It is accepted the same way: we get a success response and a session id
   straight away.

3. Check the status of that longer request. Within about a minute it stops.
   The agent has already replied with its opening message ("I'll start by
   creating a brand kit, then draft the blueprint"), but the session then
   reports that it failed, with progress still at zero — so it never began
   rendering at all.

4. Look for a reason. There isn't one anywhere: the video status says only
   "failed", the error field is empty, and none of the status responses
   include a failure code or message.

5. Repeat step 2 a few times. It fails roughly half the attempts rather than
   every time, and the failure rate got noticeably worse over a three-hour
   period. Short prompts kept working throughout.

The clearest example we saw: one long prompt completed successfully at 01:36,
and the identical prompt — same text, same account, nothing changed — failed
at 03:00 the same morning.

Output

CLI version

Using heygen rest apis only

OS

mac0S 26.5.1