The Agent Economy
Upcoming wave of Agentic AI and its potential impact
Personal opinion and thoughts based on my observations and experience/ will continue to evolve/update this
Part II: https://medium.com/@gopivikranth/the-agent-economy-part-ii-is-it-time-for-enterprises-to-own-their-intelligence-layer-with-a7a6d2ddf590 || Part III: https://medium.com/@gopivikranth/the-agent-economy-part-2-a43b64f5003c
We are at the inception of a massive wave of Agentic AI which will change the way existing software and analytics operate in the industry. Agentic AI driven workflows will reimagine major portions of SaaS, legacy software, analytics products and services. A few factors are at play
- Cost to write software and number of people required is plummeting, and speed to create new applications is increasing
- Agentic AI and agents even in their current form are effective at taking up tasks and LLMs reasoning capabilities are increasing at a rapid pace (a job is a bundle of tasks)
- Existing software is bundled and bloated with features which may or may not be needed and are expensive — This will unbundle — 1000’s of new start ups will hit the problem at various angles along with internal teams that will try to cut the cord on current SaaS
- Cost of LLMs is getting cheaper and smaller models are effective to perform single tasks (even on device)
Given these Agentic AI provides an enormous opportunity to reimagine a range of saas products and services (specially in data and analytics products) in the next few years. Many areas which were cordoned off with well established products and platforms are up for grabs, Initially these will be specific tasks that can be done cheaper faster and more effectively with agentic AI and going take 4–5 form factors.
** will be adding examples for each as I come across products or startups
1. Adapters: These essentially leverage a LLM plus a specialized data to tune it to high efficacy to perform a task. It will be good at the said task but might not be useful for other areas. These would be fast to construct and can be provided as microservices or APIs. These are also likely to use smaller models and can be done on device (ex: apple intelligence adapters).
- A similar situation is likely when certain tasks need to be done on prem platforms without any calls going to larger online LLMs since the data or the information is sensitive in nature. These will also be more cost effective
- There are also a range of adapters that can be very useful at a consumer level which do specific tasks for me as an individual and are offered on a pay as you go basis if not run locally for a one time fee on a permanent basis.
2. Agentic Insights: Today largely LLMs are given a prompt, and they will output a range of information as well as artifacts (like an email, image etc.). however, the synthesis capabilities of agents are impressive and they can be programmed to reason over a range of data sources to provide a usable insight — That can improve the quality/and or speed of the decision at hand. (ex: your product is impacted by regulatory changes, in which case an agent can keep a watch on various regulatory web sites, data and information sources globally and provide the insight on when a change occurs, what the change is and potentially what that might impact).
- In these cases specific agents are employed for specific problems which are providing insight as an output to the users. These could also be achieved by specific models
3. Agentic workflows: Every enterprise has specific strategies and processes that they are implementing which differentiate them to some degree from their competitors. This is in addition to the domain in which they operate in. There are hundreds if not thousands of such workflows in each enterprise depending on their side, and an even larger number of workflows when viewed across an industry space. This is likely the largest area where agentic AI will have massive impact. Many of these workflows will be redone with agentic Ai making them more automated, efficient and more secure.
- This will be one or more AI agents, which are imprinted with the know-how of the specific process and are connected into a workflow to execute the tasks within the enterprise. This will be with some human in loop supervision initially. These agents will also be highly configurable and have mechanism to easily ingest domain and process related knowledge (no code)
- There will also be specialized agents (or sophisticated adapters) which are part of the workflows which capture the expertise of one or more people for a certain process (for example a support experts knowledge can be imprinted to an agent which can function as a expert model based on his prior cases, writings, data and his interactions). This can be further extended with the ability to connect to other services which can get additional expert information on a case-by-case basis.
4. Agentic Apps: New application layer which are a combination of a set of workflows, agents and adapters packaged together or orchestrated by an LLM to function as an application. These are smaller, nimbler than traditional SaaS and likely a lot of current SaaS applications will have to be agentified to be competitive. (On the other hand, this is hard as it will effect current revenue models and not many would want to cannibalize their existing revenue) Which in turn leads to the gradual cut the cord effect.
5. Autonomous function specific collection of AI agents: This is the long-term promise and might take a few years to play out. If we were to take an example of digital marketing in the future this would be a combination of marketing co-pilots (coming from big tech) + army of Marketing AI agents/agentic workflows, a few agentic apps that can fully recreate or supplement a digital marketing teams full capabilities. Bringing together all the individual agents together and optimizing across channels for a desired objective will produce the marketing plan and necessary execution. Even this bringing together will be done by a managerial agent leading to a team of agents in the future delivering the capability of a specific sub function /function much faster, cheaper, better and at a larger scale vastly improving the benefit delivered. This can function with human in loop or depending on the specific business function and scale either in supervised or in fully autonomous mode
- *References: As I find industry articles, videos, references and examples of tools and new startups, I will continuously add them. Feel free to share any reference you find and I will include them
- 1. July 23: https://www.youtube.com/watch?v=Vy3OkbtUa5k || Meta Zuckerberg — Zuckerberg predicts a world with billions of personalized AI agents and emphasizes the importance of open source development. These agents will execute several tasks/activities and empower both users and enterprises
- Adapters apple intelligence for on device: https://developer.apple.com/videos/play/wwdc2024/102/?time=95
3. NIMS: https://www.youtube.com/watch?v=pKXDVsWZmUU || at 44 mins || at 50 mins
- Assembling teams of Nims instead of writing software code:
- Applications will be assembled by connecting different specialized Nims as a team
- some Nims will be reasoning agents that break down tasks, plan missions, and orchestrate other Nims
- Other Nims will retrieve information, use tools like optimization algorithms, run queries etc.
- A “team leader” Nim will coordinate the team, break down subtasks, reason over outputs
- Humans assemble and coordinate teams, so in the future companies will have collections of Nims that can be combined modularly into applications by assembling their capabilities, not coding.
- Contrasted with traditional software written with instructions, future apps will emerge from dynamically composed teams of pre-trained AI models/agents with different specializations.
- Reasoning agents will flexibly strategize and direct activities of their specialist Nim teams based on the problem, continually adapting and re-planning as needed.
- Mirrors how human teams operate — with a planner/manager coordinating diverse experts, rather than defining everything up front.
- Programmers will focus on developing competencies of individual Nim agents, not defining full system logic and workflows.
5. https://youtu.be/8Pfa8kPjUio?feature=shared // (Minutes 43–45) -Jensen NVIDIA CEO — LLM reasoning
6. https://www.youtube.com/watch?v=g7jwu-QCoNE || Servicenow ceo on agentic workflows
