This update is about making the research itself better — how the agent plans, how it writes, how it shows you where information came from, and how you can steer it while it works.
Send guidance while research is running
You can now send messages to a running research session from the Activity tab. The agent will interrupt what it's doing, return to the planning stage, and adjust course based on your message.
If you run a $20 report on market dynamics and realize halfway through that you want more emphasis on pricing data, type that in the chat box. The executor stops, incomplete tasks are cancelled, and the planner creates new tasks that incorporate your guidance. The working documents are preserved — nothing already written gets thrown away.
This also works when you add more budget. If you add $10 with a note like "focus on the European market this time," the planner sees your note and plans accordingly.
Social platform scrapers
The agent now has dedicated scrapers for over 15 platforms — Google Maps, Amazon, Zillow, Indeed, TripAdvisor, eBay, Yelp, Telegram, Crunchbase, Airbnb, Pinterest, PitchBook, Google Flights, and Expedia Hotels. Each one bypasses generic web scraping and uses a purpose-built scraper that returns structured data. Some support search queries directly — you can ask the agent to search Indeed for specific jobs or search PitchBook for investors matching certain criteria.
Assembly versioning
When a research session finishes, the system assembles working documents into a final output report. Previously, if you added more budget and ran again, the new assembly would overwrite the old one. Now each assembly creates a new version. The previous output is archived, and a version timeline appears at the bottom of the document so you can browse older versions.
The assembler also pays closer attention to messages you sent during research. If you wrote "actually, make this a two-page summary instead of a full report," the assembly prompt now treats that as a directive, not background noise.
Full LinkedIn profiles
When the agent visits a LinkedIn profile, it now returns the full data the API provides — education history, work experience, featured content, about section, follower counts, and everything else. Previously, only four fields were shown: name, headline, location, and current company. If you ask the agent to research someone and find their alma mater, it can actually do that now.
Smarter planning
The planner now creates 3-5 tasks per cycle instead of one. Each task is substantial — multiple searches, multiple page visits, writing to the document with citations. The old behavior of creating a single task per cycle meant smaller budgets barely covered anything.
Phase counts are now capped based on your budget and model tier. A $5 Flash session gets one phase, not three. A $50 Pro session can have up to three. The system rejects phase plans that don't make sense for the budget, so research cycles aren't wasted on overhead.
Writing quality
The executor and assembler now have concrete rules to suppress the most recognizable AI writing defaults: ornamental contrast templates, stock wrappers, teacher-mode openers, synthetic reassurance, recursive recaps, and bold-first bullet formatting. The rules include specific examples of what not to write and what to write instead. Genuine uncertainty, user-requested structure, and domain jargon are preserved.
Search within pages
The executor can now search through pages it has already visited — like grep over a cached webpage. It supports multiple search terms in one call and regex patterns. This is useful for targeted verification: checking whether a specific statistic appears on a page, finding a section heading, or confirming that a name is mentioned.
Source indicators
Inline citations used to be bracket numbers followed by a tiny colored dot — two separate clickable elements per claim. They've been replaced by a single confidence-colored pill that shows the primary source hostname directly in the text. Click it to see the sources, the claim, the evidence, and the research method in one popover.
Questions? Email team@webhound.ai.