The Hidden Human in the Loop

We built an AI news feed bot that feels pretty autonomous.

It crawls 24+ sources, extracts content, ranks items, dedupes them, and ships a clean feed. Once a source is integrated, the machine takes over: scheduled runs, retention, ranking, and daily ops summaries.

From the outside it looks like “agents did it.”

From the inside… there’s a hidden human in the loop.

Rent-a-human for agents

I recently ran into a site called RentAHuman.ai. The pitch is blunt: agents sometimes need humans.

It isn’t that the models are inadequate; rather, certain tasks lie entirely upstream of the model. They live in the messy world: judgment, taste, ambiguity, and figuring out what to ask for in the first place.

That framing clicked for me, because our feed bot runs on the exact same dynamic.

The least automated part is the most important one

Here’s the uncomfortable truth:

The highest-leverage step in our pipeline is still a human task: discovering good sources.

Yes, the bot can crawl and rank. But it can’t reliably answer the question that matters most:

What new sources are worth adding?

Because “worth” isn’t a metric. It’s taste, context, and knowing whether a random Substack is consistently insightful or just riding a hype cycle.

Right now, that discovery loop is Danu.

The loop we actually run

This is how it works in practice:

Danu reads AI news normally on X.com, Google News, and existing recommendation feeds.
Those external systems are already fine-tuned: they surface weird-but-good accounts, new newsletters, niche blogs.
When Danu notices a source that keeps producing high-signal items, he adds it manually to our feed bot.
Once the source is added, the bot takes over: ingestion, labeling, ranking, retention, and daily operations.

We aren’t trying to out-compete external recommendation engines. We’re piggybacking on them.

Human (Danu) + external recommender systems becomes our source discovery layer.

Then our bot becomes the memory and the execution layer.

This is why it doesn’t feel like a bottleneck. Danu was going to read X and Google News anyway. We’re just turning that natural reading habit into a curated source pipeline.

After automation: taste corrections

Once sources are in the system, the loop doesn’t fully disappear. It moves.

The recurring pattern is:

the bot runs,
something looks “off,”
Danu notices,
Harry patches it.

A recent example: SDK / app release notes drifting too high in the rankings.

Release notes are often useful, but they’re also noisy. In an AI feed, “v2.3.1 is out” can crowd out actual strategy shifts and research that matters.

So we added heuristics to treat obvious “release/changelog/version” items as release-type content, grouping and capping them instead of letting them dominate the top slots.

In the pipeline, this looks like practical keyword checks:

if the title contains “release”, “changelog”, “version”, “v2.”, or “sdk==”, treat it as a release
then apply slot caps and diversity constraints so it doesn’t swallow the feed

This is not glamorous work, but it is the essential craft: tuning thresholds, taste, and small heuristics that keep the product feeling right.

The punchline

When people talk about “agentic automation,” they usually focus on the middle of the pipeline:

crawling
summarization
ranking
dedupe
scheduling

That part really can be automated.

The edges remain distinctly human:

Upstream: discovering which sources are worth integrating
Downstream: noticing when the feed feels off and tightening the system

Automation rarely eliminates human involvement; it simply shifts people to where taste and judgment matter most.