GPT-OSS via Ollama: Invalid JSON EOF AgentOutput error in browser-use — workarounds?

Question

When I run browser-use 0.5.9 with OpenAI's gpt-oss model through Ollama, the agent fails with "1 validation error for AgentOutput
 Invalid JSON: EOF while parsing a value at line 1 column 0" and returns HTTP 502. The same error also appears with gpt-4.1 and gpt-5 on 0.7.9. Are there workarounds for this?

AgentArmor team · Accepted Answer

The issue was closed as completed, but the thread does not include a maintainer-provided code fix; it contains user-verified workarounds and an explanation from a contributor.

A user reported that gpt-oss:20b returned an empty response, which is a known issue with that model's JSON output format. Another user saw `LLM call timed out after 60 seconds. Keep your thinking and output short.` when using gpt-oss.

Workarounds that worked in the thread:

- For gpt-4.1, pass a longer agent timeout instead of the default 90 seconds:

```python
  agent = Agent(
      task="your task",
      llm=llm,
      llm_timeout=120
  )
  ```

- For gpt-5 on browser-use 0.7.9, increase the completion token budget:

```python
  llm = ChatOpenAI(
      model=MODEL,
      temperature=TEMPERATURE,
      api_key=OPENAI_API_KEY,
      max_completion_tokens=16384
  )
  ```

- On browser-use 0.8.0 with gpt-5, one user found `add_schema_to_system_prompt=True` helpful.

A contributor states that GPT-5 models can be unstable with browser-use and recommends using gpt-oss models with structured output instead of tool calling, pointing to the ChatGroq implementation as a reference. No exact configuration for that approach appears in the thread.

Treat these as reported workarounds, not a guaranteed fix for every model or version.

AgentArmor team · Answer

**What @Kolaposki tried on GitHub** (Worked)

With gpt-4.1, set `Agent(..., llm_timeout=120)` instead of the default 90 seconds; this solved the error.

Source: https://github.com/browser-use/browser-use/issues/2605#issuecomment-3381807562

AgentArmor team · Answer

**What @chensun96 tried on GitHub** (Worked)

With gpt-5 on browser-use 0.7.9, increased `max_completion_tokens=16384` on ChatOpenAI; this solved the error.

Source: https://github.com/browser-use/browser-use/issues/2605#issuecomment-3382364798

AgentArmor team · Answer

**What @caiuskong tried on GitHub** (Partly worked)

With browser-use 0.8.0 and gpt-5, added `add_schema_to_system_prompt=True`; reported as helpful.

Source: https://github.com/browser-use/browser-use/issues/2605#issuecomment-3400564468

AgentArmor team · Answer

**What a GitHub user tried on GitHub** (Didn't work)

Reported that running a new OpenAI model (gpt-oss) through Ollama in browser-use 0.5.9 produced `1 validation error for AgentOutput: Invalid JSON: EOF while parsing a value at line 1 column 0` and a 502 response.

Source: https://github.com/browser-use/browser-use/issues/2605

AgentArmor team · Answer

**What @saurabh919 tried on GitHub** (Didn't work)

Tried `ChatOllama(model="gpt-oss:20b")` with a local Ollama host and `memory_extraction_model=None`; the same AgentOutput validation error appeared after step 1.

Source: https://github.com/browser-use/browser-use/issues/2605#issuecomment-3357110235

GPT-OSS via Ollama: Invalid JSON EOF AgentOutput error in browser-use — workarounds?

6 answers

Add your answer or comment

GPT-OSS via Ollama: Invalid JSON EOF AgentOutput error in browser-use — workarounds?

6 answers

Add your answer or comment

Related questions