Research Engine — an autonomous research producer

The always-on system had a consumer, a daily and weekly briefer that summarises new research for me, but no autonomous producer feeding it. With nothing generating fresh material, the briefer could only re-surface things I had already commissioned and read, which is low value. The Research Engine is the missing front end: it decides what to research and pulls the trigger, on a schedule, with no human in the loop.

The interesting problem here is not "can an agent do research." The producers already exist and are proven. The problem is doing it autonomously without generating what I think of as workslop: plausible-looking research nobody asked for and nobody needs, produced on a cron because the schedule said so.

What it does

It runs three distinct jobs off one watchlist I maintain by hand. News answers what is happening now and produces a digest per active topic per day. Progression answers what is changing in a domain and runs weekly against a recency gate. Foundations answers what the bedrock of a domain is, works through a curriculum I write, and goes quiet when it has covered it, which is also the corpus the tutor build teaches from. Each run is routed to the cheapest producer that can do the job well. Over time it steers on lightweight good-or-bad feedback rather than per-file approval, because at this volume reviewing every file would defeat the point.

How it avoids producing noise

The design lesson worth keeping is that roughly seventy percent of the quality control is deterministic, enforced in code that the selection prompt cannot talk its way past. Eligibility, novelty checks against what already exists, volume caps, and a substance floor all run as hard gates. Only one step is a genuine judgement call: picking and ranking which fresh topics are worth researching today. Everything around that single soft step is mechanical.

That split is what makes autonomous production safe. The one place the agent exercises taste is wrapped in a deterministic envelope, the usefulness judgement is pre-loaded into the watchlist I maintain, the briefer is the audit surface, and my feedback is the steering signal. I do not review each file. The volume self-regulates: on a quiet day with no genuinely new signal, it produces fewer files rather than manufacturing filler to hit a quota.

Where it actually is

It has been running daily since mid-July across twenty-four active topics, with another eight paused, and has landed 338 news digests on twenty-eight separate days between the fourteenth of July and today. Those digests are also the source for the public news page on this site, so the output is not only filed, it is published.

The part I will not overclaim is reliability. Three days in that window produced nothing at all, and on one of them every scheduled routine I run failed to land output the same morning. Nothing alerted me. I found the gap days later by reading a report that happened to look back far enough, which is a worse failure than a loud crash, because a scheduled job that quietly produces nothing looks exactly like a quiet day with no signal. The engine's own design principle, that a quiet day should yield less rather than filler, is precisely what makes its silence hard to distinguish from success.

It is deliberately an orchestration layer and nothing more. It builds no new researcher. It decides what is worth knowing and when, and hands the actual work to tools that already do it well, which is usually where the real payoff in an agentic system sits.