Webhound

A research engine for your agent

Artificial Intelligence

Search

Visit WebsiteSee on Product HuntTwitter ⧉

depth dial is a really usefull touch 👏 agent research infra heating up fast

"Both answers can look finished" whether you spent ten minutes or ten hours is the real problem this is addressing, and it's a genuine one. Curious how Webhound communicates the confidence level of the output to the user, like does a $2 report come with any signal about what it probably missed or what claims are still weakly supported, or does it always look equally authoritative regardless of how much budget was consumed?

Taking the blunt-feedback request at face value, because we shipped the same stopping problem in a different domain and got it wrong in a way worth warning you about.

We run an AI that makes real phone calls — booking, chasing refunds, following up. Same shape as your research agent: no natural finish line, and "it stopped" and "it succeeded" look identical from the outside.

On the dollar budget: I think it's a genuinely good control and a bad stopping rule, and it's worth being clear which one you're selling. A budget bounds your downside — real, valuable, and the thing that makes the pricing honest. What it can't do is tell you whether the answer is done. Those come apart in the case that actually hurts: budget consumed, report returned, and the question was answerable in the first ten cents but the agent spent the dollar circling. The user can't tell that from a finished-looking report, and the budget was never designed to catch it.

The failure that cost us most was worse than that, though. We had chains that terminated and reported success when the real outcome was that every attempt had been rejected before it ever went out. Nothing crashed. The model wrote a confident completion. The gap was that we let the model self-report the outcome.

What fixed it, and it's the bit I'd offer you: the agent emits a structured decision — done, continue, stop-and-ask, exhausted — never prose, and then a separate validator checks that decision against ground truth and can override it. The model proposes; something dumber and more literal disposes. We also had to make "stopped because the world said no" a first-class terminal state distinct from "stopped because I ran out" — a number that says do-not-call-again is a permanent stop, and retrying it is worse than useless.

Mapping to yours: budget-exhausted and question-answered should be visibly different outcomes in the returned report, not both rendered as a finished document. If I can't tell which one I got, I can't tell whether to spend another dollar.

On your second question — sources and notes do help me decide what to trust, with a caveat we learned the hard way: citation covers the trigger, not the claim. A report can cite entirely real sources and still assert something those sources don't support. If your research notes show the reasoning step and not just the inputs, that's the more valuable half.

What does your report look like today when the budget runs out mid-question?

A nice direction, especially the cited reports angle. One thing that would make this more useful for me as a user: a way to pause a long-running agent and resume later without losing the work it has already done, plus an export to plain markdown or a Notion page so I can drop the output straight into my team's knowledge base. Right now I worry about kicking off a multi-hour run and having it vanish if I close my laptop.

It would be really helpful to have a way to export the agent's reasoning trace or intermediate findings, not just the final report. Sometimes I want to dig into how the agent arrived at a particular conclusion, especially when I'm using it for due diligence work where I need to show my work.

The budget dial is the right primitive. We maintain a pricing dataset where the answers change monthly and every number needs a source you can defend - and the failure mode of most research tools is a confident summary over dead or circular citations. If the reports really keep the working documents attached, that's the part worth paying for: auditing the claim afterwards, not just reading it.

The budget dial is a smart framing. Research has no natural finish line, and looking finished is exactly how a shallow answer slips through.

The thing I would want in the output is a fetch date on every source, so a report from last month can be re-checked instead of trusted forever. Does the dataset mode keep a timestamp per row?

makes sense, flagging the lack of evidence as a limitation instead of just going quiet is the right default. thanks for the answer

The budget-based depth control is an interesting way to make research effort explicit. How does Webhound decide when a claim needs further verification, and can users inspect why the agent stopped researching a particular lead?

Moe — the disagreement-surfacing answer to Gal is solid. My research problem's different though: federal contract award data and past-performance records aren't scattered across the open web, they're stuck behind a clunky government portal with no public API or real search. Does Webhound handle sources like that, or is it built for open-web research?

The dollar budget is a clever constraint, but the stronger idea might be making research depth explicit. Most tools hide the stopping decision behind a confident-looking paragraph. Here, at least, I can decide whether a question deserves five dollars or five minutes. I like that “not enough evidence” can be a valid output.

Budget as the stopping primitive answers how much, and there's a second question sitting under it: how much does the same $5 vary? Two runs on one question at one budget follow different leads, and an agent that follows leads is path-dependent by construction — whichever source it happens to open early reshapes everything after it. The report is a function of the budget and of which door it went through first.

That lands hardest exactly where Clemente was pointing, on the MCP path. A human feels a thin answer and re-runs it. An agent takes the first report as ground truth. I do eval work on my own app's generated output, and the number that changed how I ship wasn't the average score — it was the spread across identical inputs. The mean looked healthy for weeks while the bottom of the distribution was quietly unusable.

Have you measured that spread on a fixed question and budget? And does the per-claim confidence score reflect run-to-run stability, or only the evidence inside the single run that produced it?

Depth over speed is a refreshing pitch when everything else is racing to answer in two seconds. Exposing budget as the control on research quality is smarter than hiding it behind a vague quality slider. When it builds a dataset rather than a report, how does it handle two sources that contradict each other, does the row keep both values or does the agent pick one?

This looks useful, I lose hours copying company details off websites into a spreadsheet by hand. Having the choice between a clean dataset and a fully cited report covers pretty much every research job that lands on my desk. If I ran the same query again next month, would it give me a fresh dataset I could diff against the old one to see what changed?

I'd love to see how Webhound handles edge cases, like sources with paywalls or outdated information. How do the research agents adapt to these challenges?

I see that this is positioned as the "research engine" behind your AI agent, but it wasn't clear to me how this is different than getting the frontier model to do deep research and continually prompting it to continue researching after the default stopping point.

probe real submit pathBudget as the stopping rule makes sense, especially when another agent is waiting on the answer. The trust layer I’d want is a short handoff note: what was checked, what was intentionally skipped, and which unresolved claims could change the decision if someone spends another hour.