Building a Research Agent with Gemini 3 + Deep Agents

0:00Hey, this is Lance from Lang. Gemini 3 just came out today. This was a highly anticipated model release. And looking at the results here, we can see it's extremely strong across the board. And there's this wide battery V valves comparing Gemini 3 Pro, 225, Sonnet 45, and GPD51 with the bold indicating state-of-the-art performance. So, Gemini is doing extremely well across the board. But I do want to call your attention to a few specific evals that are particularly relevant for building agents and that's these three right here. Vending bench 2, Terminal Bench 2 and Sierra Tow Squared Bench. Now vending bench 2 is for long horizon planning tasks. Terminal bench 2 is for

0:42terminal based largely coding tasks. There's around 100 of them. So these are tasks you can actually execute using a command line in a terminal. Now ta squaredbench is gentic task in the context of real world like customer support and so one of the things that Gemini emphasizes in the release is that it's very strong for building agents and I want to test that and I'll test it using our deep agent harness. So deep agents is an open-source agent harness we released and the nice thing about it is there's a few built-in tools that are very general and very useful. notably a planning tool, a sub aent delegation tool and simple file

1:22system tools for reading writing files, glob grip and using a shell. So these are built-in tools to our agent harness and you can of course provide your own tools as well as instructions. Now I recently created this quick starts repo. It has quick start for deep research because that's such a popular task and I've added support for Gemini 3 to this repo. So let's have a look at that. So here's a UI we can use to interact with a locally running deep

1:49agent deployment that is using Gemini 3. And let's just look at this first and I'll explain how it works in the repo briefly. Go ahead and submit our research request. So Gemini is going to use the native tools provided to do various things. We can look first. It's going to call this to-dos tool to create

2:06a few different tasks. We can see the tasks listed here. They're all pending. Here is the first task to basically just write the research request to a file. We can see it kicks off this sub agent. Sub agent does some research for us. We can see it's all saved here. And we can see the final research report is written to this

2:24final report.md. Nice with some citations as well. Now what's happening here is we're seeing Gemini 3 being able to successfully do a few different things. Plan using this to-dos tool that we give it. Write files. Kick off sub aents. This is a sub agent that it calls to actually kick off a research task. It's able to then analyze those results, update its to-dos, noting completion, write a final report, and give us kind of a summary of its findings as well as providing this final report. Now, these files we're writing are actually just in the Langraph state object, and so they live in memory currently. But with DHS,

3:04you can actually supply a backend that would write to a fi to write to a file system. So it's really testing the ability of the agent to call various file manipulation tools. Read file, write file that we provide. And the back end just happens to be the langraph state object. Now if we zoom back to the quick start repo, you can open up this folder. That's where the readme for everything we just did is shown. And this actually covers how to get started and what I did. I made sure that my Google API key was set. That's all I needed here. And I can run this in two

3:34different ways in an interactive Python notebook as well as a Langraph deployment which is exactly what we just saw. I kicked off a langraph deployment simply running this command and I start up this UI. You can see the instructions here and that's what we were just looking at this deep agents UI interface. Now open up the notebook I can show you the prompts and the code in a little bit more detail. So first this just the task deep research for research. We supply two different tools. Tableau search tool and a think tool. You can find these in the deep agent folder. Just open up this

4:06subfolder and you can see this tools.py. Those are your tools. Load them in. And instructions. So these are just instructions for the research task. You can modify these any way you want. Of course, these are some instructions that I found to be useful with some heristics that I like for steering agents. But of course, these are just prompts. And I visualize the prompts in notebook. You can go through them on your own time if you're interested. And custom sub agents

4:33provide here. Now, here's where specifically we initialize Gemini 3. So again, it's pretty simple. You're just going to import from here. Lang chain Google Genai right here. Set the model to Gemini 3 Pro preview. There, set the model. Now, this is just create a deep agent. Again, this is using our deep agents harness. I supply my model. I supply any custom tools. For example, for the research task here, I supply the instructions we just talked about very briefly above and I supply any custom sub agents. Now, in the notebook, I can interact with this agent in the same way we did in the UI, but in this case,

5:08we're directly using the deep agents package, which we import. This is built on top of Langraph. We can invoke it like any other LN graph graph just like this. Pass in our messages with the research request. And you'll see the messages here for display in the notebook are actually quite similar to what we saw in the UI. Here's the human input. Here is the tool call to write to-dos. Great. We write the research request to a file. Great. And that's all

5:33based on instructions. So in the prompt, we tell the agent, hey, write the request to file so you can reference it later if you need. Nice. We update the to-dos. And we call our custom research sub agent. That's great. We give it a topic to research. And this is the result of all that research that happened within the sub agent. So that's the beauty of using sub aent to isolate context. It'll do a bunch of work and just return the results back to the parent agent. It gives us a nice deep dive on contact

6:03engineering with a bunch of sources. That's great. We update to-dos again. We write our final report based upon what the agent provided. Fantastic. Update our to-dos. We read the research request. Good. We mark all our to-dos as correct. Verify that we satisfy the original request. That's great. And we're done. We can even fetch the file from the Lang state object just like this. We can view

6:31it. And here's our final report. And all this is logged to Lenus. We can open up the trace. Click here. And this gives us an overview of everything that happened in granular detail. And we can see we are calling Gemini 3 Pro preview. So great, we're using the Gemini 3. Now, here's really the key point when working in the notebook or using our local deployment of our research agent with Gemini 3. We can see that it can use

6:56these various deep agent native tools. We can see that it uses these native tools, notably file manipulation tools, planning, sub aent delegation very effectively. This looks really good. The reports generated look quite nice. Token usage is moderate as expected for research tasks, but it's not unreasonable. And you can easily modify this for your own use case. So Gemini 3 looks looks very promising. Of course, the benchmarks are fantastic. It seems

7:24to be a great model for building agents. And we can see with our own deep agent harness, it's able to utilize the various tools really effectively for task like research. I encourage you to have a look at this. Feel free to play with it using the deep agent harness. And all this will be checked in. Again, you can check see our deep agent quick starts repo and see the deep research readme for all the instructions about how to run this yourself. So hopefully this is a useful overview of both Gemini 3 and using it with our with deep agents

7:46for a popular task like research. Thanks.