AI agents are all the hype now. Startups are raising millions of dollars with the promise of AI agents. As my friend Sumanth pointed out, in an ideal scenario, “you run an agent once and forget about it…it works continuously and forever on whatever it has been assigned to do”. According to Sumanth, “agents will also change competitive dynamics in a way that I suspect most people haven’t woken up to. There will only be one winner in each job to be done…no silver medal as each agent has infinite capacity to service job requests…so the best agent in each situation will win all the time.” He calls it the “Highlander” principle-”There can only be one”
But what exactly are agents? How can we apply these agents to our own domains.
If you break down the hype, agents are pretty simple.
Agents are prompts.
Agents have access to tools.
Agents are controlled with an orchestration layer.
Before the age of AI and LLMs, we didnt have prompts. Instead we would write large programs with a lot of if else conditions and rule engines. And the tools sometime became part of the rules leading to a lot of spaghetti code.
Instead of prompts and tools we had microservices. Microservices also needed an orchestration layer, hence that is available in both the approaches.
With the advent of AI, the access to tools is driven by the AI. This leads to more flexible systems where we dont need to update our code to support new process of new customers. Instead, if our prompt is good enough, the AI will handle it. Or, in the worst case scenario we just need to update our prompt.
Now lets see where we can create differentiation.
Prompts and AI models are a commodity now. So there is very little you can do in this layer.
Orchestration layers are also becoming a commodity with tools like LlamaIndex or Langchain or one of the many no code low code tools cropping up taking care of the orchestration.
So, your main differentiation boils down to your tools.
And this is the main reason we at Ozonetel have been able to deploy agents in the CX space so fast. We already have well defined tools for all the processes in CX. Since are a platform first, we already have well defined APIs and tools. So we built a drag and drop orchestration flow designer which integrates with our tools and hooked it up to a prompting engine. This allows our engineers to cook up agents on the fly.
So, if you want to take advantage of agents, you can adapt the following pipeline:
- Identify common repetitive tasks in your process.
- Create tools(functions, APIs) that encapsulate these tasks.
- Experiment with prompts to see if the AI is able to detect when to call what tool.
- Deploy with an orchestration engine.
AI Agents in CX was originally published in oneCXi on Medium, where people are continuing the conversation by highlighting and responding to this story.
