Chatbots, Copilots, Agents and AI Workers
TL;DR
- LLMs provide technology to implement core skills
- Core skills + UX + communications = GenAI apps
- Chatbots, copilots, AI agents and AI workers are different types of GenaAI apps
- GanAI apps can share the same core skills but provide different flavor of UX and communications
- ...different packaging of core skills
Consider the existing types of GenAI applications. One particularly useful way to understand all the different types of GenAI applications is to look at them as the set of “core skills” that are packaged in different ways. A given chatbot, copilot, AI agent or AI worker can all be based on the same set of “core skills”, i.e. perform the same business task or a job - but they do it differently in terms of UX and external communications. In other words, the same set of “core skills” is packaged differently…
Packaging
The term packaging comes from the automotive industry. Most auto manufacturers use the same engine in multiple vehicles (in multiple “applications”). An internal combustion engine consists of the main block and its internals as well as many additional, auxiliary components such as cooling systems, piping, alternator, heaters, A/C compressor, oil and fuel pumps, etc. Each vehicle that shares the same engine will have the same main block but its auxiliary components will be arranged, i.e. packaged, differently in each vehicle due to constraints of each individual application (sedan vs. truck vs SUV). 🚗
The analogy with the GenAI applications is apt. While a specific chatbot, copilot, AI agent or AI worker can share the same “core skills”, the many auxiliary components and capabilities will be different, meaning that their shared “core skills” are packaged differently.
Example
Let’s look at this packaging in an example. Imagine that we are building a GenAI parcel shipping functionality. At a high level, we can design it as a set of multi-modal AI agents or tools:
- AI agent for integrating with UPS
- AI agent for integrating with FedEx
- AI agent that creates PDF with shipping label
- AI agent that interacts with the user to collect all necessary information
- The main orchestrating AI agent that manages the workflow across all other agents
This multi-agent application can use any arbitrary set of open-source libraries, LLM models, commercial products, and programming languages to develop a modern agentic implementation. It can be deployed on the cloud or on prem.
Now, regardless of how it was developed and deployed, this agentic implementation of the parcel shipping “core skills” can be packaged as:
- A chatbot that acts as a single natural language communication channel and all operations are happening on the background per user’s request. Custom or integrated chatbots focus on a single communication channel and work mostly in a text-based request-response mode.
- A copilot that is integrated, for instance, into the user's inventory system. When the user clicks on the “Ship” button in its system, the internal workflow kicks in and makes the REST call to the integrated copilot. This, of course, requires a custom integration between “core skills” and the host application but provides a more seamless and feature-rich user experience.
- An AI agent that acts similarly to a micro-service in a large mesh of micro-services or other agents. In general, AI agent can be viewed as a classic micro-service where compute logic is replaced with LLM. The “core skills” in such scenarios are accessible to other agents or micro-services via API protocol.
- An AI worker, a hirable AI agent, that can be hired and becomes part of the existing human team. It is named Josh, for example, and anyone in the company can email or Slack Josh to ship parcels. Josh communicates with everyone using Email, Slack or other means to gather all the necessary information, creates and sends back the shipping label to use.
As you can see from this simple example, the same GenAI implementation, i.e. “core skills”, can be packaged differently. At Humatron we expect that many builders of domain-specific vertical GenAI applications will take advantage of these different packaging strategies when developing their products to serve a wider range of potential customers.
On the other side, there is a clear need for different packaging from customers' point of view - given the same “core skills”, some need a simple chatbot, others will need a fully integrated copilot, still others will require a hirable AI worker implementation.
Humatron and AI Workers
Packaging for an AI worker is probably one of the most complicated ones.
An AI worker is an AI agent that can be hired. That entails a long list of capabilities and functionality that this AI worker must support to become interviewable and hirable, be able to blend into existing human team, organizational structure and culture, use the same communication channels that are used throughout the organization, support human curation and autonomous work, etc.
Humatron API platform provides many of these capabilities out-of-the-box through our Worker API and Platform API reducing the development efforts from months to days.