Adept has published a technical walkthrough of the Adept Workflow Language (AWL), the proprietary authoring format the company says powers its enterprise agent product. According to the Adept blog post, AWL is a syntactic subset of JavaScript ES6 that allows users to compose multimodal web interactions on top of Adept’s models.
The company defines AI agents as “software that can translate user intent into actions,” and describes AWL as the mechanism for doing that at production scale.
What AWL is
AWL sits between natural language and executable code. Users write workflows — sequences of instructions the agent will follow — using a mix of function calls and plain English. The post describes several AWL functions: click() takes a label like click("Compose"), which causes the model to locate the target on screen and generate a bounding-box coordinate call (clickBox("Compose", [52, 75, 634, 979])) that the actuation layer executes as a synthetic web click.
The act() function takes natural language input and invokes what the post calls an agent reasoning loop. Commands passed to act() serve as the agent’s task prompt; the agent then makes a step-by-step plan and issues the same types of commands recursively.
The post presents three versions of a workflow that accomplish the same task. The leftmost uses only functional AWL calls — goToURL() and click() — giving the model no latitude. The middle version mixes functional calls with act(). The rightmost is written entirely in natural language, leaving the agent to make its own judgments about how to proceed. The post frames this as a spectrum: more prescriptive workflows trade flexibility for reliability; more open-ended ones give the agent room to adapt when interfaces change.
What Adept says the agent can do
The post includes several examples of end-to-end workflows. One reads a PDF of event attendees, extracts the first person’s details, and creates a new lead record in HubSpot. The post then shows the same workflow adapted for Salesforce: according to the post, the only changes needed were the URL and a few field names specific to Salesforce — a process the post describes as taking under five minutes.
A screenshot from what the post calls Adept’s internal workflow viewer shows the agent’s intermediate state during the HubSpot workflow. The post describes the agent reflecting on its last action, assessing the current screen state, and confirming its next step — including which element to click. The post notes that “the agent’s thinking here is not hard-coded: the agent loop invoked by act() means that our agent dynamically assesses — at inference time — the outcomes of previous steps.”
Additional examples listed in the post include: extracting patient information from a PDF and updating a record in an Electronic Medical Records system; pulling data from a Google Sheet to create a customer record and invoice in Stripe; triaging inbound emails and updating lead status in Salesforce; and extracting contact details from a contract PDF before searching a state database for an active license. The post notes that the last workflow involves the agent solving a CAPTCHA, “because it was instructed to do so.”
The agent’s design properties
Adept describes its agent as engineered for three properties. The first is reliability: the agent “can easily be kept ‘on rails’ to consistently execute a workflow.” The second is robustness: the agent “is resilient to changes in its execution environment, and can successfully carry on despite these variations.” The third is ease of authoring: instructions “are quick and simple to write, and can even be a few lines of natural language.”
The post argues that the multimodal foundation — the model perceives the screen directly rather than relying on API integrations or HTML parsing — is what makes the robustness property possible. The same underlying capability that allows the agent to understand a HubSpot form also allows it to work with tools like Bing Maps, which the post describes as using a 2D canvas that is “too difficult to manipulate programmatically” using conventional approaches.
What Adept says the approach enables
The post lists three categories of downstream effect the company says AWL enables: lower barriers to automation authoring, access to more complex workflows involving unstructured data, and greater resilience in existing automated processes. The post says that because the agent reasons at inference time, its automations are “dynamic and responsive to underlying web changes,” which the post frames as complementary to existing robotic process automation tooling.
The post does not provide benchmark numbers, error rates, or customer counts. The reliability figure cited — “greater than 95% reliability” for enterprise customers who work directly with Adept — appears in the earlier post about Adept Experiments, not in this one.