Agent Completions gaps today: plan around streaming and PDFs

Sume's Agent Completions page lists what is not available yet: non-image attachments (images are the only supported type), streaming and a synchronous OpenAI-style choices[] response, continuing a prior thread with thread_id and assistant turns, and team-owned threads. If your GPT-6.1 Sol or Sonnet 5.5 agent expects any of those, design around them now instead of finding out in production.

The list and the workaround

This is taken from the "Not available yet" section of the Agent Completions docs. Workarounds below are design choices, not features; nothing in them depends on unreleased behavior.

Completion is pushed, progress is not

A finished run can be pushed. A signed POST goes to your public HTTPS communication.webhook_url with up to ten attempts and a ten second timeout, as described in Run webhooks. Progress is not pushed, so a long task needs polling in the meantime. Keep the webhook as the fast path and a poll as the backstop.

Design rules

• Make each run self-contained: the instruction, the data in input, the cap and the schema. Assume no memory.
• Persist your own state between runs, keyed by your ids, and pass only what is needed.
• Bind output_schema when a program will read the result, per Structured output.
• Keep the cap per run; a pipeline of runs gets a cap on each, plus a total you track yourself.
• Read finished media from the durable URLs in output rather than rendering inline.

When to revisit

Check the page again before building anything that assumes a gap closed; this list is the contract, and it can change. For work that needs conversation, a Format or an interactive chat in the app may fit better, and Jobs and results shows how to read what runs produced. Keep API keys and signed URLs out of logs, per Safe automation.