Deep Research jobs are getting stuck in_progress and never finishing

shanegary · March 31, 2026, 6:21pm

I have multiple Deep Research jobs stuck with status=in_progress. we are passing the multiple hour mark. Same prompt works in the chat interface. Not sure what to do. This is an amazing service, I really don’t want to have to write my own research agent. Any ideas?

Mustan_lokhand · March 31, 2026, 7:51pm

Hi
Could you provide more details about this

What model are you using ?
Code snippet if possible

patrick864 · March 31, 2026, 8:01pm

I am having the exact same issue all day, almost 100 queries have been stuck in progress

Mustan_lokhand · March 31, 2026, 9:05pm

Looking into this , will keep you posted

shanegary · April 1, 2026, 6:08pm

Code snippet — run_deep_research() from workers/services/llm_client.py:

from google import genai

Agent and polling constants

_DEEP_RESEARCH_AGENT = “deep-research-pro-preview-12-2025”
_DEFAULT_POLL_INTERVAL = 10.0 # seconds between polls
_DEFAULT_MAX_WAIT = 3600.0 # 1 hour timeout

async def run_deep_research(
prompt: str,
*,
poll_interval_seconds: float = _DEFAULT_POLL_INTERVAL,
max_wait_seconds: float = _DEFAULT_MAX_WAIT,
resume_interaction_id: str | None = None,
) → DeepResearchResult:
client = genai.Client() # uses GOOGLE_API_KEY from env

  # --- Create or resume interaction ---
  if resume_interaction_id:
      interaction_id = resume_interaction_id
  else:
      interaction = await client.aio.interactions.create(
          agent=_DEEP_RESEARCH_AGENT,
          input=prompt,
          background=True,
          store=True,
      )
      interaction_id = interaction.id

  # --- Poll until completed, failed, or timeout ---
  start = time.monotonic()
  last_status = None

  while True:
      elapsed = time.monotonic() - start
      if elapsed >= max_wait_seconds:
          raise LLMServiceError(
              f"Deep research timed out after {elapsed:.0f}s: "
              f"id={interaction_id} last_status={last_status}"
          )

      interaction = await client.aio.interactions.get(interaction_id)
      status = getattr(interaction, "status", None)

      if status != last_status:
          logger.info("Deep research poll: id=%s status=%s elapsed=%.0fs",
                      interaction_id, status, elapsed)
          last_status = status

      if status == "completed":
          break
      if status in ("failed", "cancelled"):
          raise LLMServiceError(f"Deep research {status}: {interaction_id}")

      await asyncio.sleep(poll_interval_seconds)

  # --- Extract results ---
  text = _extract_interaction_text(interaction)
  sources = _extract_search_sources(interaction)
  return DeepResearchResult(
      text=text,
      search_sources=sources,
      interaction_id=interaction_id,
  )

Observed behavior from worker logs:

Interaction is created successfully (gets an interaction_id)
Polling loop runs, status never transitions to “completed”
Agent model: deep-research-pro-preview-12-2025
SDK: google-genai (Python async client via client.aio.interactions)
API calls: client.aio.interactions.create() then client.aio.interactions.get() in a loop

I will dm you the prompt and the run ID.

shanegary · April 1, 2026, 6:09pm

actually I don’t see how to dm you

lukashaas · May 13, 2026, 10:09pm

Following up here - please let us know if you are still running into these issues with the latest Deep Research agents.

Selvakumar_Eswaran_C · May 27, 2026, 3:15pm

I am still facing this issue on the Gemini 3.5 Flash model

ehostrun · June 1, 2026, 11:35am

Gemini Deep Research Interactions stuck in in_progress after May 26 schema migration

Hello Google Gemini API team,

We are seeing a regression with Gemini Deep Research via the Interactions API after the May 26, 2026 migration to the new steps schema.

We run a production Telegram bot that takes a Polymarket event, builds a prompt, starts Gemini Deep Research, then delivers the final report as PDF to the user.

Since May 26, some Deep Research interactions never complete. They remain in:

status: in_progress
output_text: empty
error: None
steps: only one user_input step


They stay like this for hours. There is no completed, failed, requires_action, cancelled, or expired terminal status.

**Environment:**

- Python 3.12.3 (production), Python 3.10 (test instance)
- google-genai==2.7.0 (production), google-genai==1.75.0 (test instance)
- httpx==0.28.1
- Model: deep-research-pro-preview-12-2025 (production), also tested deep-research-preview-04-2026 and deep-research-max-preview-04-2026
- The issue reproduces on both SDK versions

**API method in production bot:**

```python
client.interactions.create(
    input=full_prompt,
    agent="deep-research-pro-preview-12-2025",
    background=True,
    store=True,
)

Then we poll with:

client.interactions.get(interaction_id)

What we observe:

Before May 26, we successfully ran 300+ Deep Research tasks across different prompts and Polymarket events.
After May 26, some tasks get stuck indefinitely.
For stuck tasks, interactions.get(id) repeatedly returns in_progress, with no output and no error.
The only step is user_input; there are no model/tool/search/url steps, so it looks like the agent worker never starts after the interaction is accepted.
Network tests from the same server are stable. Raw HTTP requests to generativelanguage.googleapis.com work consistently.
The issue is not the SDK client lifecycle bug: our real code keeps the genai.Client instance alive.
Deep Research works normally through the Gemini web interface (AI Studio) during the same time period when API calls are stuck.
A separate CLI implementation using the same server, same model, same universal prompt, and same events has completed 8/10 runs successfully when using:

client.interactions.create(
    input=prompt_text,
    agent="deep-research-pro-preview-12-2025",
    background=True,
    stream=True,
)

This suggests the background=True, store=True polling path may be using a different execution path or queue than the streaming path.

One interaction completed successfully (2.3 min) using a brand-new API key from a fresh Google Cloud project with store=True. Subsequent requests with the same key and identical parameters returned to the stuck behavior. This suggests possible per-key or per-project rate limiting or queue starvation, rather than a pure code-path bug.

We also tested / considered:

The issue does not look like prompt parsing failure, because the stuck interaction never proceeds past the initial user_input step.
It does not look like collaborative planning waiting for approval, because there is no model_output plan step and no requires_action status.
The new steps schema parsing works for completed interactions.
The problem reproduces across all three available Deep Research models (deep-research-pro-preview-12-2025, deep-research-preview-04-2026, deep-research-max-preview-04-2026).
The problem reproduces with trivial prompts (e.g., “What is 2+2?”), not just large production prompts.
The problem is that some interactions never reach any terminal state.

Stuck interaction IDs:

# Production bot (deep-research-pro-preview-12-2025, without store=True):
v1_ChdURWtaYW83S0Z0aUY0LUVQNnFTQ21BWRIXVEVrWmFvN0tGd........
  stuck 31min, killed by bot restart

# Test bot (deep-research-pro-preview-12-2025, with store=True):
v1_ChdRVmtaYXR6N1BLX0FnOFVQMHFxSTRBMBIXUVZrWm.....
  stuck 16+min

# Test script (deep-research-max-preview-04-2026, store=True, trivial "2+2" prompt):
v1_ChdUWGNaYXJiZk04X2JqdU1QNzZTRHNBZxIXVFhjWmFyYmZNO......
  stuck 10+min

# Test script (deep-research-preview-04-2026, store=True) — SUCCEEDED:
v1_Chc5SDhaYXFMMEx2bkJqdU1QOVlTd3lRaxIXOUg4WmFxTDBMdm5CanV.....
  completed in 2.3min ✓

# Test bot (deep-research-preview-04-2026, store=True, new API key from new project):
v1_ChdDb1VaYXVQSEs2bmJqdU1QZ2J1QmdBZxIXQ29VWmF1U......
  stuck 13+min

Questions:

Is there a known issue with Deep Research Interactions created with background=True, store=True after the May 26 migration?
Are background=True, store=True and background=True, stream=True routed through different backend execution paths or queues?
Should production integrations use stream=True first and then resume/poll with interactions.get(), instead of pure polling?
What is the expected terminal behavior for a Deep Research interaction that fails to start? Should it eventually become failed or expired?
Can you confirm the current default for agent_config.collaborative_planning for Deep Research?
Should we explicitly set:

agent_config={
    "type": "deep-research",
    "thinking_summaries": "auto",
    "collaborative_planning": False,
}

Is service_tier relevant for Deep Research reliability? Could flex cause accepted interactions to remain in in_progress indefinitely?

Expected behavior:

If the task is accepted, it should either start producing agent steps and eventually complete, or fail/expire with a terminal status.
It should not stay in in_progress indefinitely with only user_input.

This regression is affecting paid user-facing workflows, so any guidance from the Gemini API / Deep Research team would be very helpful.

Thank you.

Topic		Replies	Views
Gemini Deep Research Interactions stuck in in_progress after May 26 schema migration Gemini API deep-research	2	245	June 16, 2026
Deep Research API (deep-research-pro-preview-12-2025) — jobs stay in_progress indefinitely with complex prompts Gemini API prompt	6	260	July 31, 2026
Deep Research API - Interaction stuck "in_progress" for some domains Gemini API bug , api , gemini-api , deep-research	32	694	April 3, 2026
Deep Research Agent is `in_progress` for >1h (potentially indefinitively) Gemini API deep-research	3	130	May 13, 2026
Regression: Deep Research Max via Interactions API fails with HTTP 500 after in_progress, while standard Deep Research works Gemini API gemini-api , deep-research	4	154	June 22, 2026

Deep Research jobs are getting stuck in_progress and never finishing

Agent and polling constants

Related topics