Agent SDK verbatim_prompts vs Sume Agent Completions input

If you build a video agent on the Claude Agent SDK for Python and forward customer text, verbatim_prompts (default False) makes user messages bypass @path expansion and slash-command dispatch. Sume's Agent Completions solves a similar problem differently: the input field is data, written to a file, not instructions.

What does verbatim_prompts do?

Per the SDK releases page, v0.2.158 added verbatim_prompts, default False. With it on, user messages skip @path expansion and slash-command dispatch.

What does Agent Completions do with input?

POST /v1/agent/completions takes exactly one of instruction or messages[]; assistant turns are rejected. The separate input object is data and is written to /workspace/inputs/sume-action-input.json. generation_spend_cap_usd is required and has no default, and model is only sume-agent.

Which problem each solves

They guard different layers.

• verbatim_prompts stops your own client from rewriting text a user typed.
• input keeps variable data out of the instruction text.
• Neither replaces a spend cap. In the SDK the cap is on the Sume call; in Agent Completions it is the required generation_spend_cap_usd.

When to use which?

Use the SDK when your agent runs in your process and needs your tools. Use Agent Completions when you want Sume to run the run and send a signed terminal event to a webhook. See Agent Completions.