In the artificial intelligence space as of recently, there have been a concerning number of buzzwords which heavily exaggerate or mask the reality of computer program functionality. Words like "agentic" can often seem like unparalleled magic to those less familiar with programming and large language model technology. The truth, in this case, is that an "agentic" program is fundamentally a standard, deterministic computer program that provides a message to an LLM in such a way that it will output a structured response which is recognized by the program, and then used to execute a command or task.
For example, a user might receive an email, and they might use a program which calls an LLM to autonomously decide whether to create a support ticket based on the email's contents, or do nothing. An example of the message sent to the LLM by the program would be:
Read this customer email and return JSON only.
Email: "Hi, I was charged twice for my subscription this month. Can someone fix this?"
Allowed actions:
- "CREATE_BILLING_TICKET"
- "CREATE_TECH_SUPPORT_TICKET"
- "IGNORE"
Return format: { "action": "", "priority": "", "department": "" }
The LLM, based on the above input message, would output the following:
{
"action": "CREATE_BILLING_TICKET",
"priority": "medium",
"department": "billing"
}
The program would then execute:
if (response.action === "CREATE_BILLING_TICKET") {
createTicket({
department: response.department,
priority: response.priority
});
}
The program here is a basic "agent." LLMs are a necessary part of this flow because they handle the non-deterministic interpretation step before deterministic program execution. Conventional programs are predictable because they follow defined rules, fixed logic, and exact instructions. Given the same input and state, they will produce the same result every single time. This is what is meant by deterministic. They, however, fundamentally have no way of understanding natural, messy, variable human language. LLMs, at a high level, are the opposite: they generate outputs by calculating which words or tokens are most likely to come next, based on statistical patterns extracted from massive amounts of text. Unlike normal programs, they are not following fixed hand-written rules for every case. In the agentic flow outlined above, the LLM decides what the user likely means, and writes a response in a format that the deterministic program is designed to understand. Then, the deterministic program validates that decision and executes it.
There's nothing inherently wrong with the word "agent" or "agentic," we use it ourselves, but it often confuses people and makes a simple reality seem like mysterious magic, since it's not often explained clearly. With it now explained, we have been developing a Python-based agentic program called CoResearch, designed to conduct LLM-driven deep internet research on a specialized topic, and write full-length scientific research papers based on the LLM's thought process as well as the discovered materials, in a standard format with full, verifiable citations.
In one instance, the CoResearch agent identified a plausible, testable correlation between two recent breakthrough research papers, “Advancing regulatory variant effect prediction with AlphaGenome” and “Prime editing-installed suppressor tRNAs for disease-agnostic genome editing.” The correlation proposed was a testable bridge between the two papers: PERT's ability to rescue premature stop codons may depend heavily on AlphaGenome-predicted mutant mRNA availability.
In simple terms, the paper on prime editing-installed suppressor tRNAs shows a way to install suppressor tRNAs that can read through premature stop codons and restore protein production, a method the paper calls PERT (prime editing-mediated readthrough of premature termination codons). But that rescue mechanism can only work if the damaged transcript still exists in the cell long enough to be translated. The AlphaGenome paper, meanwhile, predicts how variants affect expression, RNA coverage, splicing, chromatin accessibility, and other regulatory features. If that correlation holds, AlphaGenome-like models could serve as a pre-treatment filter, flagging which nonsense mutations are most likely to respond well to PERT before attempting treatment.
To be clear, this hypothesis may very well prove to be false. That does not make the process useless. In science, a failed but well-structured hypothesis can still be valuable because it narrows the search space, clarifies what should be tested next, and gives researchers a more precise question to investigate. The important part is not that the model instantly produced a confirmed discovery, but that it generated a plausible, testable direction from two complex papers that otherwise may not have been directly connected. This is precisely the kind of connection LLM-driven research agents can be useful for.
We are hoping to use this agent to generate new testable scientific hypotheses and potential correlations, published on Auric Research to benefit humanity one small step at a time.