Summary of work • I read the agent-sdk repository and the examples directory, and prepared clear steps to run the SDK in headless (non-interactive) mode, using GPT-5-codex, with the conversation running inside a Docker sandbox via DockerWorkspace. How to run headless with GPT-5-codex and DockerWorkspace 1. Prerequisites • Docker installed and daemon running. • API key available: ◦ OpenAI direct: export OPENAI_API_KEY=... ◦ LiteLLM proxy: export LITELLM_API_KEY=... and have a proxy base_url. 2. Setup • From repo root: make build 3. Model and environment choice • OpenAI direct: ◦ LLM model: openai/gpt-5-codex ◦ Env: OPENAI_API_KEY • LiteLLM proxy: ◦ LLM model: litellm_proxy/openai/gpt-5-codex ◦ Env: LITELLM_API_KEY ◦ base_url: your LiteLLM proxy URL 4. Minimal headless flow (what the script does) • Detect platform: linux/arm64 on Apple Silicon, else linux/amd64. • Create LLM with the GPT-5-codex model and your key. • Start DockerWorkspace with: ◦ server_image="ghcr.io/all-hands-ai/agent-server:latest-python" ◦ host_port=None (auto-pick) ◦ platform=detect_platform() ◦ forward_env=["OPENAI_API_KEY"] or ["LITELLM_API_KEY"] ◦ mount_dir=os.getcwd() (so files like FACTS.txt persist on your host) ◦ extra_ports=False (headless) • Create agent via get_default_agent(llm=..., cli_mode=True) to disable browser tools. • Create a Conversation with the workspace, send your task, run, close. 5. Running • Put the minimal script above in a file (e.g., run_headless_docker.py) and: ◦ uv run python run_headless_docker.py • Or run inline with a heredoc via uv run python - <<'PY' ... if preferred. Notes and gotchas • Headless behavior: cli_mode=True and extra_ports=False keep it non-interactive. • Persistence: mount_dir=os.getcwd() ensures files created inside the container appear on the host. • Platform: Apple Silicon should use linux/arm64; others linux/amd64. • Docker must be available and running; ports must be free (auto-picked if host_port=None). • GPT-5 family is automatically routed through the OpenAI Responses API and encrypted reasoning is enabled by the SDK, no extra config needed. Examples suitable for this use case • Best match: examples/02_remote_agent_server/02_convo_with_docker_sandboxed_server.py ◦ Adapt its LLM block to model="openai/gpt-5-codex" (or litellm_proxy/openai/gpt-5-codex) and forward the matching env var. This runs a DockerWorkspace conversation non-interactively. • Equivalent pattern: examples/remote_execution_and_sandbox/2_remote_convo_with_docker_sandboxed_server.py • GPT-5 behavior reference: examples/01_standalone_sdk/23_responses_reasoning.py ◦ Shows Responses API behavior for GPT-5 models (not Docker-based), useful for understanding how the SDK routes GPT-5. _{View full conversation}