Webhound - A research engine for your agent
by•
Research has no natural finish line. An agent can spend ten minutes or ten hours on the same question, and both answers can look finished. Webhound lets you choose how much work the question deserves. Give it a question and a dollar budget. It follows leads and checks weak claims until the budget is consumed, then returns a cited report or sourced dataset with the sources and working documents behind it. Run Webhound yourself or call it from your agent through MCP or the API.
Replies
Interesting! That framing of budget-as-stopping-primitive is definitely worth attention. Most research tools stop when the prose sounds finished rather than when the work is actually done. Curious whether you guys plan the search tree up front and rank leads by expected payoff? Is it greedy step by step and just stops when the meter hits zero?
Webhound
@artstavenka1 Webhound plans coverage first, then revises the plan as sources reveal new leads. The executor works each batch, and the verifier can reopen weak or unsupported claims. The budget controls the run without locking Webhound into its initial plan or a greedy path.
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.
Webhound
@rich_sun It is different, but not entirely, especially now with /goal and the likes-- you are right to make the association. I would say some of the biggest differences stem from Webhound being a multi agent system where the chain of thought, artifacts, and memory are designed around web research.
That being said, I also personally don't want to be constantly prompting my frontier model to dig deeper-- I want to be able to tell it how much compute to use and have it come back once it has finished, not earlier.
If a $5 run hits a dead end early, does the user get to adjust the budget mid-stream or is it strictly set at the start?
Webhound
@freya_jensen_d Yes, you can increase the budget while Webhound is researching. You can also stop the run early if you no longer think the question deserves more work.
Having said that, we’ve seen many examples of runs where the fourth or fifth lead got past what first looked like a dead end.
How does Webhound handle conflicting sources? If two credible references disagree on a fact does the report flag that discrepancy or try to weight one more heavily based on the research trail?
Webhound
@aiden_pearce7 Webhound does both, depending on the evidence. It looks at where each claim originated, how current the information is, and whether primary records support it. If that gives Webhound a clear reason to trust one source more, the report uses that source and explains the choice. It will also flag the existence of the conflicting source in the report, with reasoning on why it chose not to use it.
If the evidence remains split, Webhound flags the discrepancy instead of forcing a conclusion. You can open the claim to see both sources, the evidence from each, and Webhound’s confidence in its assessment.
More budget gives it more time to look at as many possible conflicting sources as possible, and as a result more transparent coverage of the topic.
Congrats on the launch! qq When the budget runs out mid-run, does the report tell me what it did not get to?
Webhound
@alieksia Yes. When the budget runs out, Webhound finishes the report with the evidence it gathered and lists the gaps it did not reach or could not resolve.
You can inspect the research plan and working documents to see which branches remain open. Webhound also suggests follow-up questions aimed at filling those gaps. If one matters, you can add more budget and continue the same run.
@mfkhalil Thanks. Over MCP, does the gap list come back as structured data or as prose in the report? Topping up the budget is a decision I would rather make in code than by reading a document.
Webhound
@alieksia The follow-up recommendations come back as structured data through webhound_get_session and webhound_get_evidence_pack. Each recommendation includes a title, description, suggested budget, and the guidance Webhound would use for the next research cycle. Your code can inspect those fields and call webhound_add_budget if the extra work passes your spending rules.
@mfkhalil
Thank you, that is exactly the shape I wanted. Having the suggested budget as a field means the spend decision can sit in code rather than in a prompt.
How well does it handle complex prompts with multiple research objectives in a single request?
Webhound
@voyager21 Very well! It turns every prompt into a research plan, breaks each objective into it's own task, and gathers evidence for each one.
The main constraint is the budget. A broad prompt with a small budget forces tradeoffs, so Webhound shows which objectives received less coverage and suggests follow-up research for the remaining gaps. You can also tell it which objectives matter most in the prompt.
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.
Webhound
@krekeltronics Thanks. Yeah I think budget it the best option we have so far, unless we can find an indicator / primitive that is more representative of how much work has been done. Also yes, that is the information that is included in the handoff, although it can be hard to identify every "unknown unknown"
ClinicFrame
Moe, blunt feedback on the core idea: the dial only works if I can calibrate it, and I cannot. I do not know whether my question is a one dollar question or a fifty dollar one, and by your own answer to Freya the fourth or fifth lead is often where the apparent dead end breaks. That means people systematically underfund exactly the questions that deserved more work.
So the output I would want is not only the report, it is what another five dollars would likely buy: how many leads are still open, how many claims are still marked uncertain, which of them the run considered load bearing. That turns the budget from a guess into a decision I can make again with information.
It matters more on the MCP path, because there the caller is an agent and nobody is sitting there to feel that the answer arrived thin. Does the run return that unfinished state as structured data, or does it only live inside the report?
Webhound
@clemente_lopez1 Appreciate the blunt feedback. I completely agree: a budget dial is not useful if choosing the number is still a guess.
This was a problem in earlier versions. Webhound now returns structured completion recommendations over MCP: what remains unresolved (if anything), what it could investigate next, and suggested follow-up budgets. It also surfaces them on the UI. The user can decide whether another $5 is worth it.
YourGPT
Congrats on the launch, Moe.
Webhound is tackling a really interesting problem with AI research the stopping problem is definitely something that gets overlooked. I especially like the idea of using a dollar budget as a research primitive instead of letting agents run indefinitely. I will defiently try this out
Webhound
@neha_8 Thanks a lot! Let me know if you have any feedback.
the budget-as-stopping-primitive idea makes sense to me. curious about the opposite failure mode though - what happens when a topic just doesn't have much written about it, like an internal process at a small company or something too new to have coverage. does it recognize early that more budget won't surface anything and stop, or does it keep spending trying to find sources that don't exist? that seems like the case where the dial wouldn't actually save you money.
Webhound
@omri_ben_shoham1 Good question. Today, Webhound continues until the budget is used unless you stop it. A few failed searches do not prove that nothing exists; we’ve seen apparent dead ends break after multiple failed approaches.
It changes tactics instead of repeating the same search. If evidence remains sparse, the report shows what it tried and flags the lack of evidence as a limitation. For private topics, you would need to give it access to the relevant internal sources.