ULTRON: Building an Autonomous AI Research Agent
ULTRON is the research agent we built when we realised chat assistants alone were not enough. A chat is great when you already know what to ask; research is what you do when you do not. This post explains how ULTRON works, what makes agent-style tools different from chat, and where they still fall short.
Chat vs. agent — the real difference
A chat assistant answers your question with what its model already knows. An agent, by contrast, plans a sequence of steps, fetches new information from the outside world, and comes back with a synthesised result. The difference is not intelligence — it is autonomy.
How ULTRON approaches a research request
When you ask ULTRON to research a topic, it goes through roughly these steps:
- Decompose the question into sub-questions.
- Draft search queries for each sub-question.
- Fetch pages, filter obvious junk, and extract clean text.
- Deduplicate overlapping sources.
- Rank the remaining sources by relevance and freshness.
- Synthesise a report with citations.
Source reliability is a design problem
The hardest part of building ULTRON was not the LLM. It was the internet. Reliability scoring — deciding which pages to trust and how much — took the most iteration. We use a mix of domain heuristics, cross-source agreement and freshness signals. It is not perfect, and we are transparent about that.
Deduplication and synthesis
Left unchecked, an agent will happily quote the same statistic from ten sites that all copied it from the same original. ULTRON collapses near-duplicate passages and prefers primary sources. When the sources disagree, it flags the disagreement instead of picking one.
Where ULTRON is useful
- Students preparing project reports and literature reviews.
- Journalists doing preliminary background on a story.
- Founders scanning a market before they commit to building.
- Analysts triaging news across many outlets.
Where ULTRON is not enough
ULTRON does not replace a domain expert. It is very good at aggregation and structure, but it will not spot the subtle mistake in a legal argument or the flaw in a medical study. Treat its reports as a strong first draft to interrogate, not a final answer.
How ULTRON connects with the rest of NO BOX RESEARCH
ULTRON is designed to plug into the other tools we build. Findings can be dropped into JARVIS for interactive Q&A, structured with EPIC for content planning, or captured by Baani AI as voice notes while you think out loud.
FAQ
Can I use ULTRON for academic research?
Yes, but always verify sources it cites, especially for anything you will submit for grading.
Does ULTRON respect robots.txt?
Yes. We only fetch content publicly available to search engines, and we honour disallow directives.
Is ULTRON open source?
Not today. Certain components may be open-sourced in future as part of our 2026 roadmap.
Related reading
How JARVIS AI Assistant Works: A Look Under the Hood
A practical deep dive into how JARVIS by NO BOX RESEARCH handles natural-language queries, retrieves knowledge and generates helpful, safe responses.
Web Scraping for AI Research: A Practical Guide
How to responsibly gather data for AI research — techniques, tools, ethics and the mistakes to avoid.