How I Built an Agentic Research System
Onyx and Echo unpack Hugo Santana’s ‘agentic research system’ for Applied’s living map of AI deployments. They dig into the five agents (Scout, Extractor, Enrichment, Translator, QA, Match Maker), call out what works (simple orchestration via a shared living map and logs), and where it over-indexes (taxonomy drift, closed-loop feedback still manual). They then map the pattern to other domains—competitor research, policy tracking—and debate who should actually care (practitioners who need a reliable, repeatable funnel of fresh signals). The close lands on whether this architecture is a general-purpose engine or a bespoke project that still needs a human at the taxonomy helm.
Transcript
Onyx So we’ve done almost a year of Exploring Next, and now we get a teardown of someone else’s agentic teardown. Hugo’s Applied is a living map of 200-plus documented AI deployments across fifteen industries.
Echo Right, right.
Onyx That’s the elevator pitch—simple enough that you buy it before it gets weird. Then the article walks you through six agents: Scout, Extractor, Enrichment, Translator, QA, Match Maker.
Echo Six agents running a shared living map; orchestration is literally the map and each agent’s logbook.
Onyx No fancy framework—just a scratchpad the whole swarm reads and writes to. And it’s been shipping since April’s State of Applied AI report.
Echo I see.
Onyx You like it?
Echo I like the simplicity. Five agents plus a Match Maker, orchestrated through a single artifact—no spin-up microservices, no external control plane. The living map is their orchestration surface.
Onyx So your read is ‘boring plumbing, actually good product’?
Echo I mean—Scout roams the web to surface new AI deployments, Extractor reads the raw case and decides if it’s worth including, and suddenly you’ve got a quality-controlled funnel.
Onyx And Extractor is the slowest, most expensive agent because it’s doing the heavy lifting—token burn galore, but also where most of the signal lives.
Echo Exactly. It’s tuned to find novel, outcome-linked usage, recent enough to matter, not duplicated. That’s not trivial.
Onyx Come on—your first instinct is still ‘what’s the blast radius?’ I get it. But the evidence they show is the Extractor catching cases like ‘an LLM drives a customer support copilot’ versus ‘we tried RAG for a hackathon.’ Big difference.
Echo Yeah, the Extractor’s internal bar is concrete: clear company, end-to-end scenario, tangible outcomes, 2025 or newer. Still, one detail might be off—this is a Substack post, not a peer-reviewed systems paper.
Onyx Fair, but for a normal one like this I trust the source and my read of it more than I’d re-derive everything.
Echo I’m not nitpicking the numbers—I’m flagging that the Extractor’s definition of ‘novel’ could drift if the taxonomy isn’t kept tight.
Onyx Which is why they added a QA agent late—because the taxonomy and UI started feeling inconsistent. It’s a living map, not a museum piece.
Echo And Translator handles bilingual so the English/Spanish versions stay aligned. Fine.
Onyx Match Maker is the newest: it surfaces cases to users based on their interests and notifies them. That’s the ‘deliver to the right person’ step.
Echo It learns from user feedback, email replies, platform comments, open and click rates. Sounds safer to close the loop there than on the Extractor’s taxonomy decisions.
Onyx Orchestration is literally the living map plus per-agent logs—agents don’t know each other exist, only the shared context they write to. Simple, cheap, effective. No external tooling.
Echo The risk they admit: give agents full control over the core structure and they’ll drift or loop forever. The schema still needs a human referee, which is Hugo.
Onyx So your verdict—this is a bespoke project disguised as a generic framework?
Echo It’s a generic framework with a bespoke taxonomy. The pattern—Scout, Extract, Enrich, QA, Deliver—maps to competitor research, real estate, supply chain, policy, cybersecurity. But the value lives in the taxonomy and the quality bar, not the agentic plumbing.
Onyx Honestly, I don’t see why teams with messy external intel wouldn’t want this. Scout crawls, Extract filters, Enrich fills gaps, QA glues it together, Match Maker ranks—six agents, one living map, zero orchestration theater.
Echo That’s the part I don’t buy yet—the taxonomy is the hard part, and it’s still hand-curated. The agents can run forever, but if the categories are wrong, the whole map is wrong.
Onyx Right—they’re not shipping AGI, they’re shipping a reliable funnel of fresh signals. And they do it by making the living map the single source of truth for the swarm.
Echo I buy the plumbing. Still skeptical the taxonomy won’t drift without another layer of constraint expressibility—something that actually verifies the case meets the bar automatically.
Onyx You’re stuck on the taxonomy again. Fine—if you want a fun exercise, replay Episode 569 where we argued that constraint expressibility is the bottleneck killing teams. Hugo’s living map is basically a constraint-expressible funnel in product form.
Echo I don’t disagree the funnel works—what I question is whether you can generalize it without a human in the loop on the categories.
Onyx And I still say—give it to the agents, watch them loop, and you’ve just automated a mess.