Adept began rolling out access to Adept Experiments on November 9, 2023, according to the company’s announcement post. The program is framed as a way to explore technology the company is developing for enterprise use cases. Each experiment is described as “a self-contained mini-tool or demo that showcases a part of our underlying tech.”
The first experiment is a workflow builder for the web that users can configure in plain language.
What the first experiment does
The post describes the Workflows experiment as a demonstration of what the company calls the foundational skill for an AI teammate: quickly learning a task from a user and reliably running it. The post gives the example of adding a lead to a CRM, which it notes “looks very different from company to company” — framing task-specificity as a key design constraint.
The post lists several example workflows the company says users have explored. A recruiter workflow advances a candidate to the next hiring stage and requests their availability from a single button click, running in a background tab. An accounting workflow opens attached invoices, extracts invoice numbers and total cost, and enters the information into accounts payable software. An insurance workflow extracts data from claims emails and inserts it into forms in separate software tools. A retail workflow, demonstrated for a new Shopify store manager, creates discount codes using a workflow trained by a more experienced colleague.
The post includes caveats. Because Workflows is an experiment, the post says it “often requires some careful prompting to make it do what you want.” On reliability, the post draws a distinction: enterprise customers who work directly with Adept to enable their workflows experience “greater than 95% reliability,” while the public experiment does not carry the same guarantee.
The model underneath
Workflows is powered by ACT-2, which the post describes as “a model fine-tuned from the Fuyu family and optimized for UI understanding, knowledge worker data comprehension, and action taking.” The post notes that in some cases the system also uses language-specific models for simple tasks like text composition.
The post describes Adept as an end-to-end multimodal agent: it “uses software just like a person would: it can perceive the screen directly via pixels and act on your computer through coordinates and keystrokes.” This approach, the post says, makes the system “infinitely extensible without the need to create hundreds of API integrations, manage user login credentials, etc.” The post gives Bing Maps as an example of a tool that uses a 2D canvas and had previously been difficult to manipulate programmatically.
Design and safety framing
The post describes several design principles reflected in the Workflows interface. When taking actions, Adept prompts the user for information needed to complete the task. It then either performs actions one at a time — for creating or testing a workflow — or auto-runs all actions with each step visible. The post states: “We share the community’s concerns about the safety of fully autonomous agents; for that reason, Adept is designed for human-in-the-loop supervision.”
Research roadmap and access
The post outlines areas the company says it is focused on: higher-level planning, improved visual reasoning, incorporating enterprise context, and learning from demonstrations.
At the time of publication, Adept described itself as already improving workflows for its first enterprise customers and said it was selecting “a small number of additional partners for 2024” for early enterprise access. New users were to be onboarded on a rolling basis. The post thanks infrastructure partners including NVIDIA, Microsoft, Oracle, and WEKA.