How this is made

Twice a day, at 8am and 8pm IST, a scheduled job on a Mac in Mumbai collects candidates from four places:

• 82 RSS feeds. Lab and company engineering blogs, individual researchers, arXiv and Hugging Face paper feeds, and a handful of newsletters. High-volume general feeds are filtered against a keyword list (agent, agentic, harness, mcp, tool use, claude code, codex, multi-agent, computer use, swe-bench, eval, context engineering, and a few more) so a general tech feed can be included without flooding the pool.
• Hacker News, above 80 points. Queried through the Algolia API for agent, agentic, coding agent, mcp, and llm. The points floor is the whole filter: it is a crude signal, and it is honest about being one.
• 15 sites with no usable feed, scraped from their index pages with a headless browser. Anthropic engineering and research, Cursor, Letta, LangChain, Zed, and others that publish without RSS.
• 12 accounts on X, read through the official X API. These are treated as a discovery channel, not as sources: when a post points at a real release or paper, the brief links the release or the paper and credits the post with via.

Every URL the pipeline has ever collected sits in a ledger, so the same story never gets served to you twice.

The split is deliberate and it has not changed since the pipeline started. Judgment goes to one model, mechanical work goes to another, and anything that can be a plain function stays a plain function.

• Python does everything deterministic. Fetching, deduplication, the seen-URL ledger, read-time arithmetic, schema validation, and file writing. If an outcome can be computed, no model is asked for an opinion about it.
• GPT-5.6-Sol executors do the mechanical passes. Every collected item gets summarized, assigned a section, scored from 0 to 10, and clustered with other coverage of the same story. This runs in parallel batches, because it is per-item work with a checkable output.
• Claude Fable 5 does the judgment. Choosing the lead, ordering by consequence, writing the lead skeleton, deciding a day was quiet, and writing the Sunday through-line. That is the editorial layer, and it is the part that gets read before it ships.

Nothing here uses an API key I typed into a config file for a vendor you cannot see. Both models run through subscription CLIs, and the costs are the subscriptions plus about 15 to 30 cents per run for X reads.

Roughly 18 to 30 items survive a collection run. The rules the curation step works under:

• One item per cluster, and the primary source beats the retelling.
• Anything scoring under 6 gets dropped unless it is uniquely important, and marketing copy is treated with suspicion.
• If fewer than three items clear the bar, the issue ships short and says so in the subject line. Padding a thin day is the fastest way for a daily to lose you.
• Every number and every quoted string in a summary has to appear in the source text the pipeline is holding. An item that cannot be grounded gets demoted to a bare link or dropped.
• Hedging is preserved. “Suggests” never becomes “shows”, and a preliminary eval never becomes a confirmed benchmark. A separate checking pass diffs the written summary against the source with exactly that failure mode named.

The web edition publishes itself. The pipeline opens a pull request against this site, automated validation runs against it, and it merges on green. A bad generation shows up as a red pull request rather than a broken page.

Email is different. Every send waits for me. I get a notification with an approve link and a hold link, and if I do not act within two hours the issue sends on its own under my name. That window is a deliberate choice: a daily that only ships when I am awake is not a daily. What it is not is an excuse. Anything that goes out is my responsibility, approved or not, and the corrections section below is how that gets settled when it goes wrong.

That review step is also what the EU AI Act calls human editorial control, which is the reason this page exists in the form it does. I would rather over-disclose than let you discover the pipeline yourself.

Every issue carries a corrections section, in the same position, whether or not there is anything in it. Most days it reads “Nothing to correct.”

When something is wrong, the correction states what the issue said, what is actually true, and links the issue being corrected. The web version of the original issue is fixed and the correction stays visible. Nothing gets quietly edited out of the archive.

If you spot something, tell me. A correction filed by a reader is the cheapest quality signal this thing has.

It does not read anything behind a paywall, so paywalled sources are marked and summarized from what is public. It does not rank by engagement, because the point of curation is to surface the good thing nobody clicked. It does not have sponsors, a paid tier, or a referral program.

And it will be wrong sometimes. A pipeline that summarizes 30 items a day at speed will conflate two similarly named labs, or read a preprint's claim more confidently than the preprint does. The structures on this page exist because that is expected, not because it is unthinkable.