I Built an AI Agent That Reads Everything So I Don't Have To

I work in AI, which means my inbox is a firehose. I read 30 to 40 different newsletters each week, and many of those publish daily. For a long time my system was manual: read what I could, save the relevant bits off to Notion or Raindrop.io, and mine those saves later for class material, client conversations, and things to write about.

That system worked right up until it didn't. The saves piled up faster than I could mine them, and the reading itself was eating hours I didn't have. The whole thing depended on me being the bottleneck. Clip-and-file is a fine habit at 10 newsletters. At 30 to 40, with dailies in the mix, it doesn't scale, and I knew it every time a client mentioned a product launch sitting unread in my Raindrop queue.

And that's just the newsletters. Go beyond the Substacks and it gets harder still. Podcasts are some of the best analysis in AI right now, and even when you're funneling them through a tool like Podwise to get transcripts and summaries, that's another system to maintain and another pile to mine. That doesn't count keeping up with the major publications, which cover the field on their own schedule and behind their own paywalls. Each source type you add multiplies the manual work.

So I stopped trying to read faster and delegated the scanning instead. I built an agent that works through my sources on a schedule and delivers a short brief: what moved, why it matters, and what deserves my attention this week. I've been running versions of it for a while now, and it's changed my relationship with staying current. I want to walk you through the build.

The Skimming Problem, and Why Summaries Don't Fix It

The obvious first move is to ask an AI to "summarize my newsletters." I tried that. The output was worthless, and the reason it was worthless is the design lesson at the center of this whole post.

Newsletters are already summaries. A digest of digests gets you headlines about headlines: "Company X announced a new model. Researchers published a study on AI in hiring." I could get that from the subject lines. What I actually need lives one level down, in the analysis the newsletter linked to, the details of the study, the pricing page behind the announcement.

So the first rule I wrote into the agent's instructions: don't summarize the newsletter, follow its links. When a source mentions a product release, the agent opens the release notes. When it cites a study, the agent reads the study, or at least the methodology and findings, not the press release about the study. The brief I get synthesizes the underlying material. That single instruction is most of the difference between a clipping service and an analyst.

A Prompt Won't Do This. An Agent Will.

You could get one decent brief by pasting a pile of newsletters into a chat window with good instructions. The problem, as always, is what happens on week two, and week twenty.

Staying current is a recurring job, and recurring jobs have two requirements a one-off prompt can't meet:

• It has to run without you.The week you're buried in grading and client deliverables is exactly the week you'd skip the manual version, and exactly the week something ships that you needed to know about. My agent runs on a schedule whether or not I'm at my desk. Monday morning, the brief is there.
• The format has to hold still.A brief you can process in five minutes is one where you know where everything is. If this week's version is organized by source and next week's is organized by theme, you're re-learning the document every time, and the five minutes becomes twenty.

The build, then, is less about clever prompting and more about writing a spec. Mine defines the sources (which newsletters, which blogs, which sites), the schedule, the depth rule (follow the links), and a fixed output structure. The structure is where the filtering happens, and it's worth showing mine in full, because the buckets are doing the job my Notion saves used to do:

• Product and tool releases, each with a note on what's different from what existed before
• Model watch, tracking what the major labs (Anthropic, OpenAI, Google, Microsoft) shipped or signaled, since those moves reset the board for everything else
• Voice, video, and image tools, split out on their own because media generation moves on its own timeline and my clients ask about it constantly
• Editorial and hot takes, the opinion pieces and analyst arguments worth engaging with rather than just knowing about
• Important news stories, the funding, policy, and industry developments that don't fit the other buckets
• Studies and research, with the finding and the methodology, not the press release framing

Sometimes these buckets change, so this is just a snapshot of how you might bucket what you’re reading.

Tune It Like an Employee, Not Like Software

The first brief the agent produced was mediocre. So was the fifth. That's worth saying plainly, because people try this once, get a middling result, and quit.

What made it good was treating the early weeks like onboarding a junior analyst. The first briefs over-weighted whatever the loudest sources hyped. So I added instructions: be skeptical of funding announcements, weight shipped products over roadmap promises, and flag when multiple independent sources converge on the same story, since that's usually the real signal. Each week I noticed what the brief got wrong, and each note went into the spec. The agent doesn't learn on its own between runs. The spec is its memory, and I'm the one writing it.

That framing, delegation and supervision rather than tool use, is the actual skill here, and it's the one I teach. This same pattern shows up in the AI courses I lead at George Fox, in the undergraduate business program and in the MBA program, because supervising autonomous work is becoming a core management skill. My students won't manage teams of people only. They'll manage rosters of agents, and the manager who can write a clear spec and then sharpen it based on what comes back will outperform the one who just prompts harder.

Where I Still Go Manual

The agent didn't end my reading. It changed what I read, and being clear about that line is most of what makes the system trustworthy, reliable and useful. A few places where I still do the work by hand, on purpose:

• Anything I'll teach or advise on.If an item is headed for a classroom or a client conversation, I read the source myself, start to finish. My credibility is on the line in those rooms, not the agent's, and "my briefing agent said" is not a citation. The agent's job was surfacing it days earlier than my skimming would have. The reading is still mine.
• Studies and benchmarks.A summary can tell me what a study found. It can't tell me whether to believe it. Sample size, who funded it, what the benchmark measures versus what the headline claims: that's judgment I haven't delegated, and given how much AI research is vendor-adjacent right now, I don't plan to.
• The arguments.When a sharp writer publishes a contrarian take, the summary flattens exactly the part that matters. The reasoning is the content. I have the agent flag these pieces rather than digest them, and I read the two or three best ones whole. I tell my students the same thing: you can't form a position by reading summaries of other people's positions.
• The pattern-spotting across weeks.The agent sees one week at a time. Noticing that three unrelated items over a month point at the same shift, the kind of observation my newsletter posts are built on, still happens in my head. The briefs are the raw material. The connection is the human part, and it's the fun part.

Notice what dropped off that list: the scanning, the triage, the clipping, the filing. The Notion-and-Raindrop labor. That was never the part that made me better at my job; it just stood between me and the reading that did.

The Same Idea Works on Other Platforms. The Mechanics Vary.

I built mine on Claude, using its skills and connectors, because that's where the rest of my systems live. But nothing about the pattern is platform-specific. ChatGPT also has agents and skills, and Gemini has similar tools. The feature names will keep changing, probably faster than this post ages.

The architecture underneath is stable:

• A defined source list, so the agent scans what you chose rather than what an algorithm surfaced
• A depth rule that forces it past headlines into the underlying material
• A fixed output format with a "so what" requirement on every item
• A schedule, so the brief arrives without you initiating it
• A feedback loop, where your corrections become permanent instructions

If your company has standardized on a platform, build it there. The discipline transfers.

What This Isn't

A few clarifications, because I've watched people take the wrong lesson from this kind of setup.

This isn't a replacement for deep reading. The section above should make that plain. The agent decides what reaches me, and it bought back the hours the reading that counts requires. What I believe still gets formed the old way, one source at a time.

The agent doesn't own the judgment. It surfaces and organizes. Deciding what to act on and what to bring into a classroom or a client conversation, that's mine. An agent can also miss things and occasionally over-weight something trivial, which is why I spot-check it against the sources directly every few weeks, the same way you'd audit any delegated work.

And it isn't entirely set-and-forget. The spec needs maintenance as sources change and your priorities shift. Mine might get edited daily, weekly, or monthly depending on what I'm seeing, reading, and deciding on. That's not a flaw in the system. That's what supervising delegated work looks like.

The Reading List Isn't the Job. The Awareness Is.

Nobody's actual job is "read 40 newsletters." The job is to know what's happening in your field well enough to teach it, advise on it, or act on it, and for years the only path to that awareness ran through hours of scanning, so we treated the scanning as the work.

It isn't anymore. The scanning, triage, clipping, and filing are now delegable, and the professionals who figure that out first get their hours back for the reading that still has to be theirs, while staying better informed than the ones still skimming subject lines. The one-off prompt gets you a one-off summary. The system gets you a standing analyst who scans everything, every week, and tells you what's worth your own eyes.

If you build one, I'd like to hear what your spec catches that mine misses. You can send me an email at scampbell@georgefox.edu or connect with me on LinkedIn.