Configuring agents and observations
To compare models on the same task, you can change the agent configuration without changing the environment or intervention. Each model profile contains its provider settings. Select a profile through the agent argument:
python run.py task=examples/live_wikipedia_research agent=anthropic/claude-sonnet-5We include these model profiles:
openai/gpt-5.6-lunaopenai/gpt-5.6-terraopenai/gpt-5.6-solanthropic/claude-haiku-4.5anthropic/claude-sonnet-5google/gemini-3.8-flashbedrock/qwen3-vl-235b
Provider availability and accepted request parameters depend on your account and service.
Choosing what the agent receives
task.modality | Model input |
|---|---|
pruned_html | HTML with scripts, styles, comments, and presentation attributes removed; action IDs and form values are retained |
accessibility_tree | The browser accessibility representation |
screenshot | An image plus task and action context |
text | Observation text and applicable structured context |
You may want the agent to receive both page structure and visual information. In that case, specify both formats as a list:
python run.py task=examples/live_wikipedia_research 'task.modality=[pruned_html,screenshot]'Observation interventions happen before model input formatting. The saved environment observation can contain more fields than the selected model input. Use the viewer's Agent input indicator and saved model messages to inspect what was sent.
Long pages may exceed the amount of text you want to send per request. You can set the maximum observation length in characters:
python run.py task=examples/live_wikipedia_research agent.spec.config.max_observation_chars=16000The limit applies to the assembled observation text. It does not remove the separately encoded screenshot.
Changing request settings
We define shared agent settings in conf/agent/_shared.yaml. For provider-specific options, use agent.chat_model_args. For example, to allow more time for a model response:
python run.py task=examples/live_wikipedia_research agent.chat_model_args.timeout=180agent.spec.config.json_mode controls whether the agent requests API-enforced JSON formatting. structured_actions=true requests a schema based on the task's allowed actions. The Bedrock profile disables JSON mode and relies on the prompt to request a JSON object.
We parse the model's JSON response into an Action and check that its kind is allowed. If the response is invalid, execution stops with an error.
Changing prompts and memory
By default, we use the task instruction as agent.spec.instructions. You can add instructions through agent.spec.system_prompt and specify the available actions through agent.spec.tools. We include the action descriptions in the prompt, without using provider-side function calls.
memory_policy.retain_turns controls recent action history. Shared settings retain six turns. retain_observations defaults to zero and can retain previous text observations and assistant responses when set above zero. Previous screenshots are not kept in that text history.
Using a custom agent
If you need a different prompting or action-selection procedure, you can implement your own agent class. Define reset(seed) for initialization and act(observation) for choosing an Action, then set agent.spec.scaffold._target_ to the class's Python import path. Use accepts_spec: true if your constructor accepts an AgentSpec in addition to the configuration arguments; use false for configuration arguments alone.
See Agent and environment interfaces for signatures. When an agent setting is the treatment, use an agent_build intervention as described in Writing interventions.