Agent · launch intelligence

Launch-intelligence agent

An autonomous agent watches a curated list of lab blogs, vendor news pages, and arXiv categories, identifies frontier-AI launches, and extracts structured fields plus a draft "what changed" note in Pere voice. Candidates land in a review queue. Approve to greenlight; reject to silence.

Agent control

11 sources watched · last scan never

Filter

Queue empty

Run a scan to populate the queue with candidates the agent surfaced.

How the agent works

Sources are curated. The agent doesn't crawl the web — it watches a small list of high-signal feeds (lab blogs, vendor newsrooms, arXiv cs.LG / cs.CL).
Extraction is conservative. The model is prompted to reject anything that isn't clearly a launch (no marketing posts, no policy commentary, no hiring news). If nothing in a source is a launch, nothing gets queued.
Numbers carry confidence flags. Confirmed (officially stated), estimate (range given), or rumour (speculative). The flag survives into the draft note.
Human in the loop. The agent never publishes. Approved candidates become the basis for a registry update or note; rejected ones get silenced so the same source post doesn't resurface.
Idempotent. Re-running the scan doesn't duplicate. Each candidate has a stable hash of its source + headline + link.