AI Agents and the Future of Business
Imagine for a moment a business in 2026 (yes, that is how quickly we believe that the business world will drastically evolve) that’s optimized by AI agents; we will call this hypothetical business “Orange Bank.” Like its competitors, Orange Bank has long had access to financial forecasting software, yet implementing AI agents has been a game changer for their business. While previous software had simply identified market trends, the new agents provided to the teams for financial analysis have effectively identified market signals from data sources that more traditional approaches may have missed - in a fully automated way. For instance, the agents retrieved signals from a regional news network overseas and deduced that a Superchips Semiconductors warehouse lost significant inventory due to significant water damage. Although the company had not yet announced this, the agent was able to let the analyst know, allowing them to adjust their estimated value for the company.
This is an important area where next-generation agents differ greatly from previous AI and analytics tools: they can identify relevant information from places that other tools and people would otherwise not due to the ability of automated systems to comb through and make sense of a greater amount of sources levering the advanced reasoning capabilities Large Language Models provide. It is unlikely that a human could navigate such vast broadcasting data in the manufacturing center overseas for Superchips Semiconductor, but the agent, noticing that a building had collapsed in the area via social media data, was prompted to investigate recent broadcasting in the region. Although the report was unrelated to the water damage, aerial footage revealed that the warehouse was compromised. The agent then compiled data from several hundred reports to estimate the impact it would have on the stock price for Superchips.
This theoretical string of events demonstrates a compelling use case for agents and how they’re able to sort through and reason with massive amounts of information from multiple sources without needing to be prompted by a user to find keywords or topics from a specific document - all due to the power of LLMs. The agent, in essence, inputs data into a language model like Meta’s Llama 2 to generate the most relevant outputs.
In fact, agents can utilize multiple language models that are fine-tuned for specific tasks along with external APIs to find information that is most relevant and accurate to the user's query, picking data with the help of one and selecting quotes from the output of another. Agents autonomously control the use of LLMs, evaluate and route results given a user input to achieve the best possible result under time or cost constraints. This opportunity could allow Orange Bank to employ a much smaller team of analysts, saving them hundreds of millions each year and improving the quality and timeliness of their decisions (which would allow them to profit in trading). Additionally, Orange Bank could more efficiently provide customer service and detailed analytics and tools into their business operations that would otherwise not have been at their disposal.
Oversight, Control Flow, and External Knowledge
So where does our product, the production-grade AI Agent framework Council, fit into this equation? While the first generation of AI agent frameworks already utilize LLMs (recursively calling them to achieve a complex task), there is no way to dictate how the LLM executes (if at all) and how much money or time is spent on such an execution. Moreover, current models can’t adapt if they do not reach a favorable result. Council and the next generation of AI technology allow greater customizability, control and optimization for users — allowing agents to better plan and work under budget constraints. Council enables users to design an AI agent with greater control flow, allowing for more reliable outputs as well as greater flexibility. This upgraded control flow will allow for better planning, resilience and budgeting, and more, resolving many of the frustrations of unwieldy AI.
For (yet another) example, consider an automated customer support bot: it can usually direct you to the resources you need but often requires more effort to feed in input than is necessary with a human. With next-generation AI, customer support bots will be much more intuitive with more automation and accuracy, more accurately perceiving subtle differences in language and providing increasingly valuable outputs to humans
A final notable benefit of greater control flow is the ability to utilize multiple LLMs, allowing an agent to choose between different language models for the best combination of outputs to produce the most useful result. Say one language model has a particularly adept multimodal functionality (allowing it to better recognize images, video, and other media better than others), while another is the best for selecting useful text, and a final model is the best at producing an intuitive response. An agent of this caliber would be able to utilize all three models for the combined best response. For example, the Agent that Orange Bank deployed, first used the multimodal model to create a transcript of what it saw, which it then sent to the second model to identify terms that may be related to water damage, that the agent finally sent to the last LLM to reason if there was indeed water damage. If at any point the agent identified an output that was not satisfactory (identified by external influence) the software would be able to recognize this and try a different language or input to receive a better result. Ultimately this is what we are striving for, a more intuitive and customizable AI agent, but also one that is robust, with predictable behavior and can be deployed at scale for products / applications