Tinyfish Launches Bigset an Open Source Multi Agent System That Builds Structured Live Datasets From Plain English Descriptions

Justy Okay, Cody, this is basically our episode four hundred fifty-six thing where I go, wait, that sounds useful, and you go, no, that sounds oversold.

Cody Yeah, because the oversell is right there in the premise. 'Describe the dataset in plain English and get a live table' is a GREAT sentence. But the actual problem is still entity discovery, source quality, schema ambiguity, and breakage over time.

Justy Right.

Cody What I do like is the article doesn't pretend it's one magic model call. It lays out a real system: schema inference with Claude Sonnet through OpenRouter, then an orchestrator using TinyFish Search, then parallel sub-agents, one entity per row, capped at six tool calls, then dedup and source links. That's a real design, not just agent glitter.

Justy I had the most normal week and somehow this still felt exciting. I spent, no joke, an hour untangling duplicate rows in a spreadsheet from two vendor exports. Then my flight got delayed and I was doing column cleanup on my phone like a tiny goblin… anyway, that is why this hits for me.

Cody That is such a bleak product-manager image. But yes, that's the strongest case for it. The argument isn't really 'AI can understand the web.' It's 'most dataset work is a repetitive pipeline, and maybe we can compress the setup cost.'

Justy Exactly. If I'm on a small team and I need, say, companies hiring for a role with location, funding stage, open req count, I do NOT want to pick a source, write selectors, normalize fields, set refreshes, and babysit it. I want a table I can inspect and export.

Cody Mm-hm.

Justy And the example they use is pretty good because it's the kind of request that lives in a meeting doc for three weeks before someone finally suffers through building it.

Cody My pushback is that the phrase 'builds structured live datasets' can hide a lot. Schema inference is nice, but natural language is messy.

Justy Sure.

Cody So the table may look clean while the semantics are fuzzy. That's the classic trap. A polished CSV can make weak recall feel authoritative.

Justy I think that's fair, but I also think you're grading it against bespoke pipeline quality, not against what people actually do now. A lot of teams are already making fuzzy tables by hand. If BigSet gives them source-linked rows and scheduled refreshes, that's already better operationally.

Cody Yeah, Justy, against manual chaos, I buy that. The source attribution matters a lot. If every row carries a link, then at least the human can spot-check instead of trusting a black box.

Justy And the refresh part is sneaky important. The article's real pitch, to me, is not one-off scraping. It's this middle layer between 'I need data' and 'I need a maintained table.' That's more valuable than a clever demo.

Cody Wait— the most convincing bit for me was actually the security design. The article gives a prompt-injection example, and their answer is pretty solid. The insert_row tool never accepts a dataset I D argument. The authorized dataset I D is captured in a JavaScript closure when the workflow starts, so the model literally never gets the capability to write elsewhere.

Justy Oh interesting.

Cody That's good engineering. The boundary lives in infrastructure, not in 'please behave' instructions. I wouldn't call the attack surface gone, because untrusted pages can still poison extraction quality, but they did think about containment in a grown-up way.

Justy That part made me trust the whole thing more, weirdly. When a system like this is honest about where the boundary should live, I assume the team has been burned before in useful ways.

Cody Nothing says maturity like having your toy break in public.

Justy No, but seriously, who should care. I think ops people, market research people, growth teams, recruiting teams, founders doing ugly early prospecting. Anyone who needs a decent live dataset fast and can tolerate some fuzziness if sources are visible.

Cody I agree with that slice. I would NOT use this, at least from what's in the article, for anything where completeness is the product. If missing ten percent of entities breaks the decision, or if field definitions have to stay stable for months, you still want a deliberately engineered pipeline.

Justy Right, so it's not replacing data engineering. It's eating the annoying middle. Which, honestly, is a huge market of internal work nobody brags about because it's too boring to brag about.

Cody And because boring work is where the money is, yes. Also, open source matters here. AGPL-three-point-zero, self-hosted on Docker, code on GitHub, bring-your-own keys through TinyFish, OpenRouter, and Clerk. That's enough concrete surface area that this isn't just article vapor.

Justy Yeah. I could be wrong, but my verdict is: not magic, probably uneven on edge cases, still genuinely practical. If it saves a team from hand-rolling ten scrapers for an internal table, that's a real win.

Cody Mine is close to that. The claim holds if you hear it as workflow compression, not truth machine. I like the architecture. I like the capability boundary. I don't fully buy the easy-language framing once the data request gets squishy, but as a system, it's more credible than most agent stuff I read.

Justy That's pretty generous from you. Should I mark the date. Cody approved an agent system… with caveats.

Cody Please include the caveats in VERY large font.

Justy Fine. BigSet: useful if you need a living table, less convincing if you need perfect ground truth. That's enough for a Wednesday.