agentic_turn_simulation_resources_server: resources_servers: agentic_turn_simulation: entrypoint: app.py domain: agent verified: false description: Whole-turn simulation for agent harnesses. The policy plays one agent turn inside an agent-harness conversation against simulated tools; an LLM judge plus code gates score the turn. value: Improve whole agent turns in agent-harness conversations (instruction following, tool use, honest reports). # Simulator and judge are separate settings; by default both use the same model server. tool_simulator_model_server: type: responses_api_models name: agentic_turn_simulation_judge_model tool_simulator_chat_completions_create_params: messages: [] max_tokens: 16384 temperature: 1.0 reasoning_effort: high tool_simulator_max_attempts: 4 tool_simulator_max_concurrency: 64 judges: - model_server: type: responses_api_models name: agentic_turn_simulation_judge_model judge_chat_completions_create_params: messages: [] max_tokens: 20000 temperature: 1.0 reasoning_effort: high judge_max_attempts: 6 judge_max_concurrency: 32 # policy | reference_turn (judge the row's recorded turn instead of the policy's). score_source: policy # Recoverable failures injected into the policy's web calls (README "Configuration"). Off: set enabled: true # to fail one early web call in row_rate of the rows. obstacle_injection: enabled: false row_rate: 0.3 kinds: {transient: 0.25, blocked_route: 0.5, empty_search: 0.25} call_kinds: [fetch, search] harnesses: null max_call_index: 2 seed: 0 # Tool simulator and judge, served with vLLM. Pass judge_base_url (one URL or a list of replicas), judge_api_key and # judge_model_name (the served model name). agentic_turn_simulation_judge_model: responses_api_models: vllm_model: entrypoint: app.py base_url: ${judge_base_url} api_key: ${judge_api_key} model: ${judge_model_name} return_token_id_information: false uses_reasoning_parser: true agentic_turn_simulation_agent: responses_api_agents: agentic_turn_simulation_agent: entrypoint: app.py resources_server: type: resources_servers name: agentic_turn_simulation_resources_server model_server: type: responses_api_models name: policy_model max_steps: 50 max_time_seconds: 7200 stop_tools: - ask_user datasets: - name: example type: example jsonl_fpath: resources_servers/agentic_turn_simulation/data/example.jsonl agentic_turn_simulation_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: agentic_turn_simulation_agent