OpenAI Agents SDK for Python 0.22

Apply the OpenAI SDK for Python 3.6 and Python 3 prompts first. These rules cover Agents SDK 0.22 agents, tools, handoffs, guardrails, sessions, approvals, MCP, tracing, and runner limits.

Agents and runner limits

• Construct agent instructions, model settings, output types, handoffs, and tools from trusted application configuration.
• Keep user, retrieved, session, and tool content separate from authoritative instructions.
• Set finite max_turns because None disables the turn limit, and set max_function_tool_concurrency because None starts all function calls emitted in a turn concurrently.
• Treat final output and typed output as untrusted until domain invariants and destination controls validate it.
• Handle MaxTurnsExceeded, model errors, cancellation, and partial streaming as non-success states.

Function tools and hosted tools

• Define narrow function-tool schemas and validate arguments independently before execution.
• Map tool names to fixed callables and never import, evaluate, or dispatch from model-produced text.
• Reauthorize each tool call against current identity, tenant, resource, and operation context.
• Restrict hosted tools, computer use, code execution, file search, and web search to approved workloads and destinations.
• Require approval for privileged, irreversible, external, costly, or destructive tool actions.
• Set a sanitized failure_error_function on function tools, Agent.as_tool(), and local MCP servers; the default formatter includes exception text in model-visible output, while None propagates errors.

Handoffs and