Thunk.AI is a software platform where you and your team create and run AI-powered work projects and processes.
This article describes the important concepts of Thunk.AI
At the heart of the Thunk.AI platform is a generative-AI model. At the moment, the default model is the GPT-4o model from OpenAI, although we expect in the near future to support a broader range of AI models.
Let's describe the concepts from the inside-out:
The platform has users. You, your teammates, and others. Each user has an AI agent
You create and work on AI-powered projects, either by yourself or with your teammates. These AI-powered projects are called "thunks".
The Thunk.AI platform hosts and runs these thunks for you
Users and their AI agents
Like any other software platform, you sign up for an account with Thunk.AI. Unlike other platforms however, each user in Thunk.AI has an AI agent.
The AI agent has access to the same platform artifacts and capabilities as the user.
The AI agent acts on behalf of the user to get work done (in Thunks).
The AI agent learns and specializes over time to become a better assistant on behalf of the user.
It is the user's AI agent that engages with the generative AI model on behalf of the user, and automates work for the user. Each user can configure the behavior of their AI agent in various ways.
Every thunk has an owner, the user who creates it, and potentially other participants, other users who take on tasks in the thunk.
What is a "thunk"?
A "thunk" is the primary artifact in the Thunk.AI platform. A thunk represents a specific AI-powered project or process or application with a goal.
A thunk has two components:
Logic: this starts with a goal or purpose, a plan (a sequence of steps) to follow to achieve the goal, and guidelines for the AI agents to adhere to while doing the work. Plans may be extremely simple or they could be complex, depending on how much direction you want to give the AI agents. As creator of a thunk, your AI agent helps you with planning.
Data or "state": this includes documents that provide context, work items with structured properties that record ongoing work, and messages (eg: emails) that carry external updates into the thunk.
A thunk has two phases of its lifecycle:
Creation phase: this is where the user and their AI agent work together to plan what should be done and how it should be done. This forms the logic of the thunk.
The execution phase: this is where the thunk is live and the project or process runs with AI agents doing work. This is the important and usually long-running phase of the thunk lifecycle. Many thunks run "forever" because they automate an important ongoing business process of a team.
The Creation phase
The creation phase is about planning. As a user creating a thunk, you could simply just provide a short goal like "I want to hire a sales lead" or "Process customer support tickets" and leave it to your AI agent to do the rest of the planning. Or you could engage in the details of the planning process along with your AI agent. How does your AI agent know how to plan? This is pretty easy actually. It utilizes the inputs available (your description of the project goal and any other relevant information). Your AI agent constructs the prompts needed to ensure that sensible planning is done for that goal, sends them off to the AI model, checks the results, and then records the results of planning in the thunk.
The outcome of the creation phase is a thunk that is ready to execute. It has a Plan template which acts like a logic recipe. For example, if the goal of the thunk is to process customer support tickets, the plan might define a workflow to follow for each support ticket. It is only a template of a plan at this stage, because each actual ticket will get a specific instance of the workflow, and the detailed execution will depend on the content of that ticket. Likewise, the data template specifies the information (fields and their types) that should be maintained for each support ticket, just like a database schema. While this sounds potentially complex, in practice, the planning phase is complete in a minute or less unless the user wants to engage in the details and modify them.
The Execution phase
Then we get to the execution phase of the thunk. Here, the Thunk.AI platform creates the live data set based on the data template, accepts data input, kicks off plan workflow tasks, and allows users and their AI agents to work on the tasks.
Where do the data inputs come from? They could directly be added by users, or they could come from AI agents that populate data (eg: in a thunk focused on hiring for an open job position, AI agents might search the web to find suitable candidate resumes). Data input may also occur via integrations with other applications (eg: a new email arriving to a user could be automatically routed to add a new data entry in a thunk).
The owner and their AI agent have some special responsibilities in managing the thunk. For example, the owner's AI agent does the work of verifying that any data entered into the thunk meets the desired criteria and organizational policies. Likewise, when new data is added, a plan-directed sequence of tasks (a "workflow") needs to run. This workflow is orchestrated by the AI agent of the thunk owner. Each task is assigned to a a specific user. The assigned user and their AI agent then work on the task. While this sounds easy, this is where most of the value of a thunk lies.
Some thunks are configured to be entirely automated. The AI agents do all the work and the human users only engage if their AI agents request intervention or clarification because of something anomalous. Other thunks can be configured to have a greater level of "human-in-the-loop" interaction.
In a separate article, we'll describe what an AI agent can and cannot do. In brief, each user's AI agent has a set of core capabilities. It can search the web, read specific web pages, it can translate text, it can read and write documents, it can read and write images. These capabilities can be easily extended by adding any number of additional "tools" that let the AI agent interact with other systems and applications. On the other hand, there are clearly some things no AI agent can do. For example, if the task requires going physically into a garden and weeding, it clearly has to be the human user doing it (however much we wish this weren't true!). On the other hand, for most online/digital tasks, the AI agent can at least help with it. Luckily for many knowledge workers, the majority of tasks fall into this category.
What does the Thunk.AI platform do?
The Thunk.AI platform hosts and runs thunks for all users. It maintains the data for every thunk and it hosts and runs the AI agent tasks for every thunk.
Since the goal is to provide a self-service platform product, Thunk.AI allows users to sign up and start experimenting with thunks without any sales engagement or human-driven provisioning. Each user even receives a small free quota of AI model usage credits to explore the system capabilities.
The Thunk.AI platform allows thunks to ingest data from other sources like Google Drive and it allows thunks to communicate with other applications via standard protocols like email, HTTPS, etc.
The default Thunk.AI service is hosted in a public cloud and is accessed, like most other modern services, through a browser tab. However, some corporate customers can choose to have a custom instance of the Thunk.AI service installed in their own cloud servers, so that they have a greater sense of control over the data in their thunks.