Sume Agent Completions request: required cap, model sume-agent

A valid request has either instruction or messages, never both, and messages may not contain assistant turns. generation_spend_cap_usd is required, and model, when you send it, must be sume-agent. The response is a receipt with an agrun_ id; poll /v1/agent-runs/{id}. Each call starts a fresh thread, so context does not carry across calls, and there is no streaming.

Keys need agent_completions:write to create and agent_completions:read to poll.

Request

curl -X POST https://api.sume.com/v1/agent/completions \
  -H "x-api-key: $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sume-agent",
    "instruction": "Make a 6-second looping product shot of a ceramic mug.",
    "generation_spend_cap_usd": 10
  }'

Common 400s

Carry context yourself

Because every call is a fresh thread, put prior facts in the instruction or pass them in messages as user turns.