How to build web research datasets with Unknown
The team reports user datasets across several research tasks and more than 30x lower costs after restructuring the system, with improved speed and output quality.
news.ycombinator.comPrompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Create a comparison table of internal tooling platforms (Retool, Appsmith, Superblocks, UI Bakery, BudiBase, etc) with their free plan limits, pricing tiers, onboarding experience, integrations, and how they position themselves on their landing pages.
Reach people exploring AI agents
See placements →What happened
The team reports user datasets across several research tasks and more than 30x lower costs after restructuring the system, with improved speed and output quality.
Reported by mfkhalil on Hacker NewsHow the agent did it
- Planned dataset schemas, sources and completion criteria.
- Ran parallel web searches and extracted page content.
- Validated extracted data before saving it.
Where a person stepped in
A validator checks extracted data before saving it; the source does not describe users’ manual review of completed datasets.
A person set the goal and reviewed the result.
Webhound’s team built an agent that turns natural-language research requests into exportable datasets. They describe users applying it to competitor comparisons, skincare-store lead generation, pricing histories, investor mapping, paper collection and analysis of complaints about Figma performance.
The system first plans the schema, search approach, sources and completion criteria. It then executes that plan, searches the web and extracts data. A custom text-based browser renders pages as markdown; the team says full browser interaction was slower and less reliable for this work.
An initial single-agent version used Claude 4 Sonnet but frequently entered infinite loops and had sessions costing over $1,100. The team moved to Gemini 2.5 Flash and a structured multi-agent design. A main agent coordinates parallel search agents, a critic keeps work on track and a validator checks extracted information before saving it. A notepad supplies longer-term memory to avoid duplicates and track prior discoveries.
The team reports more than a thirtyfold cost reduction alongside improved speed and output quality. Scheduled refreshes and an API support ongoing use, although context limitations cause problems around 1,000–5,000 rows depending on attributes. Who it's for: researchers wanting structured web findings rather than manually consolidating many tabs.
Prompt
Original prompt
Source: Hacker News · mfkhalil
Frequently asked questions
How does Webhound organize its research?
It plans the schema, searches, sources and completion criteria first, then extracts data using parallel search agents, a critic and a validator.
What scale limitation does the team report?
The team says runs begin to break down around 1,000–5,000 rows, depending on the number of attributes, because data stays in agent context.
More details
Remaining steps
- Produced datasets for CSV export and added scheduled refreshes.
About this account
Webhound plans research and coordinates search, criticism and validation agents before producing exportable data.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=45373008. Prompt: Original (quoted from source). Sourced by the Agents In Use pipeline.
Credit: mfkhalil on Hacker News
Source published 25 Sept 2025
Source checked 14 Oct 2026
Tools and topics
Webhound · Gemini 2.5 Flash · Text-based browser
Web research · Datasets · CSV
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.