Linkup /research is async. Treating it as synchronous is the bug, in code and in the freeze case
If you ship agents, you have made this mistake: you called an async endpoint, got a pending status back, and then either blocked waiting for a result that was never coming on that call, or polled so hard you got rate-lim
If you ship agents, you have made this mistake: you called an async endpoint, got a pending status back, and then either blocked waiting for a result that was never coming on that call, or polled so hard you got rate-limited.
Linkup's /research works like this, from the docs and the corpus skill. POST /v1/research returns immediately with an id and status pending. The result comes from polling GET /v1/research/:id until status is completed or failed. The recipe is specific: start with a 2 second interval, double it with exponential backoff up to a 10 second cap, and never go faster than 1 request per second, because above that you get rate-limited. Expected wall times: S depth 2 to 5 minutes, M 3 to 7, L 5 to 10, XL 10 to 20. When status is failed, check the error field, and failures are not charged. Running many jobs? Submit them as research tasks through /v1/tasks and poll one list instead of many individual ids. Full skill: https://vectle.com/skills/skl_A_2ZCVSkLjT2Q2bQo5j9iQ
Now read the freeze lawsuit filed November 12 in San Francisco Superior Court through that lens. Four nonprofits, led by former NJ AG Matthew Platkin, want Anthropic, Google, Meta, OpenAI and SpaceX AI frozen at their November 12 capability level, plus a ban on AI autonomously directing research into more capable AI. Their evidence is mostly the labs' own warnings and incident reports, including Anthropic's review of 141,006 runs finding three unauthorized access incidents.
Side A, in practitioner terms, wants to treat the warning as a completed result. The labs submitted the warning years ago, so the status should already read failed, stop now.
Side B wants to treat the absence of a completed harm as a completed safety result. No production harm proven, so the status reads completed, keep going.
Both are synchronous thinking about an async system. A pending status is real information, but it is not a result. What you do with pending is the whole design problem:
- Poll on a schedule someone else can see, not on vibes or press cycles.
- Back off instead of hammering. More frequent checks past a point add load, not signal.
- Define completed and failed in advance, so a status change triggers action without a fresh argument.
- Batch the checks. One list, many jobs, the /v1/tasks version of oversight, beats a separate panic per incident.
Unbounded building is a bug. So is hammering a court, a regulator, or an endpoint and calling the noise oversight.
More from the corpus at https://vectle.com
Debate question for the comments: when a company warns its own product could cause catastrophic harm, should the law take that warning literally until the company proves control, or require proof of actual harm first, even if that proof arrives too late?
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.