Learning Library

← Back to Library

AI Agentic Research Revolutionizes Knowledge Work

Key Points

  • AI agentic systems are rapidly transforming research across fields by automating tasks that would normally take humans hours or days, exemplified by Stanford’s multi‑agent “STORM” that produces fully‑cited Wikipedia pages in minutes.
  • Human research begins with a question and proceeds through a structured workflow: defining the objective, planning the approach, gathering data, iterating on insights, and finally delivering an answer.
  • Simple factual queries can be answered with a single search, whereas complex, multi‑step questions require reasoning, synthesis of diverse sources, and often speculative analysis—challenges that multi‑agent AI can tackle.
  • Multi‑agent research platforms replicate the human workflow by assigning specialized agents to each step and equipping them with online/offline search tools and APIs, enabling them to define objectives, create plans, collect evidence, refine findings, and generate comprehensive answers.

Full Transcript

# AI Agentic Research Revolutionizes Knowledge Work **Source:** [https://www.youtube.com/watch?v=TiNedLS_txU](https://www.youtube.com/watch?v=TiNedLS_txU) **Duration:** 00:07:35 ## Summary - AI agentic systems are rapidly transforming research across fields by automating tasks that would normally take humans hours or days, exemplified by Stanford’s multi‑agent “STORM” that produces fully‑cited Wikipedia pages in minutes. - Human research begins with a question and proceeds through a structured workflow: defining the objective, planning the approach, gathering data, iterating on insights, and finally delivering an answer. - Simple factual queries can be answered with a single search, whereas complex, multi‑step questions require reasoning, synthesis of diverse sources, and often speculative analysis—challenges that multi‑agent AI can tackle. - Multi‑agent research platforms replicate the human workflow by assigning specialized agents to each step and equipping them with online/offline search tools and APIs, enabling them to define objectives, create plans, collect evidence, refine findings, and generate comprehensive answers. ## Sections - [00:00:00](https://www.youtube.com/watch?v=TiNedLS_txU&t=0s) **AI Agents Accelerating Research** - The passage explains how multi‑agent AI systems such as Stanford’s STORM can automate the research workflow—turning simple prompts into fully annotated Wikipedia articles in minutes—by emulating the human, question‑driven process of inquiry. - [00:03:06](https://www.youtube.com/watch?v=TiNedLS_txU&t=186s) **Agentic AI Research Workflow** - The speaker explains a multi-step, multi‑agent AI process for addressing a complex, speculative legal question—whether Brown v. Board of Education applies to private schools on Mars—by defining objectives, planning, gathering data, refining insights, and generating answers. - [00:06:19](https://www.youtube.com/watch?v=TiNedLS_txU&t=379s) **Human‑AI Collaborative Research** - The passage emphasizes that future research leverages open‑source multi‑agent frameworks to augment—not replace—human scientists, enabling them to delegate routine tasks to AI and focus on high‑value innovation and decision‑making. ## Full Transcript
0:00Agentic systems continue to make waves across industries, 0:03and one of the most important use cases is a agentic research 0:08or research done by agents. 0:11And it makes sense whether you are a data scientist 0:14or an engineer or a knowledge worker in finance 0:17or medicine, or even academia, 0:21research is often time consuming, 0:24repetitive, and it involves sifting through mountains of data. 0:28And yet AI agents can accomplish in minutes 0:32what takes us humans hours or even days. 0:37So, for instance, a team 0:39at Stanford University in February 2024 0:43published on a multi-agentic system 0:46that called STORM 0:49that takes research or does research 0:53and then writes a Wikipedia page in minutes, complete with annotations. 0:57Given a simple prompt, how can I agents accomplish this? 1:01Well, first let's answer the more fundamental question 1:04how do humans actually do research? 1:07Research is one of those activities that's core to the human experience. 1:12You can't get through middle school without writing at least one 1:16research paper. 1:17And yet, the word is kind of hard to define. 1:21The word research actually comes from the Old French word 1:26researcher. 1:30Which means 1:32to seek. 1:35After 1:38or to search for. 1:43So that means that research 1:45always starts with a question. 1:50Someone seeking after the answer 1:53to a question. 1:58Oftentimes, many questions. 2:09Sometimes the question is simple 2:12and factual. 2:17Say, for instance, you're a lawyer 2:20and you want to understand, you know, what date was the case 2:24Brown vs Board of Education decided? 2:28Simple, factual. 2:29You popped into, search engine of your choice. 2:36Search 2:38engine... and outcomes and answer. 2:45May 17th, 1954. 2:50This is why we have terms such as desktop research. 2:53But more often the question is far more complex. 2:57Say that same lawyer, once in office and preparing to defend a case 3:02now has a new question such as... 3:06Does Brown versus Board of Education apply to private schools on Mars? 3:12There's no single search result for that. 3:14The answer requires reasoning, understanding 3:18of legal precedents, case law analysis, and even 3:22speculation about legal frameworks that don't yet exist. 3:26It requires a multi-step process. 3:28In response to this complex question. 3:41The first step in the process is, 3:45of course, 3:47define the objective. 3:51The second step is making a plan. 4:00Of course, 4:01after making a plan or a research plan, we need to proceed in line 4:06with that plan to gather the data. 4:12And then we need to refine our insights 4:16as we obtain them. 4:21And lastly, we need to generate an answer. 4:31This is where a agentic research comes in. 4:34AI agents, specifically multi-agent systems 4:38capable of reasoning and with access to numerous tools like custom 4:43online and offline search tools, including search APIs, 4:47can generate research findings by mimicking how humans 4:51investigate, synthesize, and iterate on knowledge. 4:56So, like the plan that a human would follow a 5:00well designed agentic research system works in the very same way 5:05it starts in response to a question 5:10by defining 5:14the research objective. 5:17And this could be the job of one agent 5:21in this multi-agentic system. 5:24Next, it would break down that plan, 5:28that original objective, into a series of steps. 5:34Or what we humans 5:35would call a research plan. 5:40That could be the job of another agent in the system. 5:44And then, of course, there would be an agent specific colleague 5:48whose goal and tools 5:51equip him to gather data. 6:00Next, there would be an agent 6:03or agents to run this out. 6:07The last two steps, which is refining 6:12based on the original steps and plan, 6:15and the insights gather and then generating an answer. 6:26Notice how similar 6:27this is to the way that humans research. 6:30It's iterative, contextual, and it builds on previous knowledge. 6:36So if you're a data scientist, developer or researcher 6:39interested in exploring your own use case in agentic research, 6:45there are many examples out there that exist. 6:48Many example implementations of multi-agent frameworks, 6:54that provide you with numerous examples. 6:58If you are a data scientist, developer, or researcher 7:01are interested in exploring your own use case in agentic research, 7:06popular open source multi-agentic frameworks exist. 7:10But just remember, the future of research isn't just 7:13humans or AI, it's humans plus AI. 7:17The real power of agentic research is in augmentation, not replacement. 7:23By offloading tedious research work to intelligence systems, 7:27you get to spend more time on high value tasks like innovation, 7:32experimentation, and decision making.