Meet the Ricochet Scientist Assistant Agent

The AI space is moving fast. Frankly, it’s a bit unsettling how quickly things are evolving and developing. 

Today, we are announcing the new Ricochet Scientist Assistant Agent (learn more at the offering landing page).

As a staff scientist, it’s hard to know where to even start. You probably like technology and tinkering with things, but who has time to dig into something that already seems overwhelming - especially when you know that by the time you get a handle on the current state of things, it’ll likely already have evolved to something completely new! 

But, as with anything complicated, sometimes it helps to back up and look at things from a first principles perspective. First principles rarely change quickly, if ever, and they are often a good place to start. 

Indeed, one of the first principles of the modern (and future) AI stack is something called “Agentic AI”. You have probably heard about OpenAI’s ChatGPT, and perhaps Google’s Gemini and Anthropic’s Claude. These are typically called “chatbots”, which are friendly “front-ends” to a much more foundational technology called a “Large Language Model” (LLM), which you can think of as the engine of the chatbot. In this analogy, visiting chatgpt.com is driving the car (and the car’s engine is the LLM under the hood). As OpenAI develops new “models” (typically a new version number), they remove the old engine from the chatgpt.com car, and add the new one in. This is a very oversimplified analogy, but it’ll serve well for this illustration.

In this blog post, we’ll be talking about our new Ricochet Scientist Assistant Agent, and, more broadly, something called “Agentic AI” - a foundational element of tomorrow’s “AI stack”.

What is the difference between an AI agent and ChatGPT or Claude? Simply put, an AI agent has the agency to perform actions in pursuit of the goal you give it. These actions can be performed over a considerable amount of time, in some cases - and will continue until the agent feels that it has satisfied its objectives. This is in stark contrast to the chat interface of chatbots like ChatGPT, which simply respond to text the user inputs into the chat interface. Over time, chatbots have been adopting more agentic-like features, with the ability to perform actions and tasks, and dive deep into research topics and sources and compile reports, but at its core, a chatbot is still simply responding to the user’s messages in a tightly coupled manner. 

Don’t get me wrong, I get a lot done with the simple chatbot interface. But it’s very limited by its nature. Human employees are much more helpful than chatbots for most complicated tasks because a human can create a plan in pursuit of an objective, execute that plan over a long period of time, assess and interpret the results, find errors or problems with the data (or original plan of attack), and then self-correct in flexible iterations of a self-updating plan, all in pursuit of the final objective. 

Interacting with an Agentic AI is much more akin to this human process than the traditional chatbot interaction. Typically, through a chat interface, you would explain, in as much detail as possible, the task and objectives, with any additional detail that might be relevant or helpful to know - just as you would explain what needs to be done to a human assistant. The agent will often reason about the task and ask any clarifying questions that it needs to begin crafting an execution plan. Based on your answers, with possibly a bit of back and forth, the agent will then craft its execution plan, and possibly request final confirmation before it begins. Upon confirmation, it starts executing the plan, step by step, which may involve performing hundreds or even thousands of search queries and visiting hundreds of web sources as it compiles the requisite information and data in pursuit of its objective. 

Today, in 2026, agents are often enlisted for research where a comprehensive report filled with synthesized data, conclusions, and citations can be reviewed by the requesting party, but this is only the first iteration of agentic AI’s eventual capabilities. 

Through something called “tool use”, agentic AI can unlock new skills that allow it to perform new and more advanced actions. Indeed, tool use is already often used during the execution of complex report compilation. The AI agent may not know how to access a particular datasource - and so it accesses a skill to learn how to perform the action. 

For example, in the course of its report compilation, the AI agent may determine that it needs to “read” a physical book that it can’t find on any free online websites. It may do a search for libraries with digital checkout options (with a PDF version of the book) and then access a skill that will allow it to sign up for a library card online with a library that offers a digital checkout of the book, check out the book, and then read the PDF to access the content. Barring that, it may find that there is an audiobook available for digital checkout and access a skill that allows it to “listen” to the audiobook to access the content and then compile the required information into its report. 

At each step of execution in its plan, the AI agent may evaluate its progress, re-evaluate and adjust its plan, and continue its execution until it reaches the end of the plan. A final review in comparison to its objectives can then be performed, and if the task has been completed, the user can be notified that the goal has been achieved. 

But looking to the future beyond simple report compilation, Agentic AI can be enlisted to perform much more complex undertakings. With the necessary data, an AI agent could perform almost any task, even those we traditionally think can only be performed by human employees. Of course, an AI agent doesn’t have a physical body to perform physical tasks, but it could easily perform actions that enlist the help of a human where needed. An AI agent can also monitor a process, including actions performed by humans, and keep track of the process, inputs, outputs, and final results, acting as a manager or owner of the overall process - and, possibly, the initial report generation. 

In a lab setting, an AI agent could also connect to lab equipment through its built-in programmatic interface (often called an “API”) to perform actions on a schedule according to an experiment’s protocol. If something goes wrong, an AI agent could access a text-to-speech skill and make a phone call to the lab technician or PI to inform them of the problem. 

The possibilities are limited only by what can be programmed into the skill the agentic AI accesses. 

Today, there are countless teams working on science-focused agentic AI tools. A quick Google search shows many mature projects that claim to replicate the abilities of a skilled scientific professional. These tools claim that these agents can perform anything from parsing a stack of PDFs into structured, queryable knowledge, to answering technical questions against the literature better than a postdoc, to running the full arc of discovery - hypothesis, experiment, analysis, and manuscript - with minimal human input. While it’s likely that some are better than others at these tasks, the writing is on the wall that agentic AI will get better and better at doing many of the tasks that staff scientists rely on human researchers to perform today. 

But, as a staff scientist who may be (desperately) looking for opportunities for efficiency gains in your day-to-day work - and with limited funds to hire more humans to help you - your only recourse up until now has been to evaluate these myriad AI assistant agent options, somehow divine which would be best for you, and then figure out how to set it up on a server so that it can actually perform work for you - all while performing your already overwhelming day job and keeping up with the latest scientific advances in your field. And worst of all, you’ll need to find time to debug issues with it, perform security and feature upgrades, and know when the technology has evolved to the point that it’s time to switch to something new. 

Seems impossible, right?

After all, the alternative doesn’t seem sustainable. Working harder and hiring more people doesn’t seem like a viable path forward. And while AI has its current limitations and risks, adoption and investment commensurate with merit, while no silver bullet, seems like a path forward that will yield compounding results with time.

Here are a few key differentiators of the Ricochet Scientist Assistant Agent

• Our team sets the agent and server environment up for you. We have extensive experience working in the National Lab setting, and can set up a purchase order / task order with procurement to ensure that everything is done by the book. We can also either set up the agent environment in the cloud, or coordinate with your Lab IT team to provision a server at your facility for the agent.
  
  
• The Ricochet Scientist Assistant Agent is unique in that it does not try to reinvent the wheel of what other scientific agent tools are already doing. Rather, our agent serves as an “orchestration layer” by pulling together other tools, agents, and skills - and then delegating tasks to each based on what they do best. This is ideal because it means that by installing our agent, you’re less exposed to technological obsolescence. Think of the Ricochet Scientist Assistant Agent like a National Lab facility itself, within which are skilled scientists and specialized equipment. When more advanced equipment becomes available, it can be acquired and can replace old, obsolete equipment. When there is a new technological breakthrough in the science, a PI who specializes in that breakthrough can be hired by the Lab.
  ‍
  Similarly, when a new AI agent, skill, or tool is developed by a team, its code can be pulled into your environment, which the Ricochet Scientist Assistant Agent can then query and delegate tasks to. The agent itself changes slowly, but new skills and capabilities can be constantly added and expanded through new tool inclusion, which our team helps facilitate for you.
  ‍
  
• When something breaks, you can rest assured that you have a team that you can call to help address the issue quickly.
  ‍
  
• Likewise, you have a team to call for training that you or your staff might need - or questions you might have.
  ‍
  
• You communicate with and delegate to your agent through a simple chat interface. While you may have some self-taught coding skills, they will not be necessary to interact with your agent. Everything is managed through a natural language interface - or even via email or text message, in some cases. 

Curious to learn more, or if your use case or preferred service can be integrated into our agent? If you’d like a demo of the capabilities of the Ricochet Scientist Assistant Agent, please reach out to hello@teamricochet.com, and we’d be happy to set up a quick intro. We also know that you might simply be curious to learn more, and we’re happy to share what we’re working on and get your feedback in a no-pressure meet-and-greet. We know that while you might not be ready just yet, an introduction can still be helpful and worthwhile. 

If anything, it’s always great to make a new connection. 

We look forward to hearing from you!