Release Notes

What changed

Image-based Post-run Checks still appear after the main automation finishes, but Cyberdesk can now start the slow model verification work earlier when the required screenshot attachment is already available.

Runs continue to show the same statuses as before: the UI switches to Running Checks only after main execution completes, then waits for any in-flight check work before choosing the final run status.

What changed

When a focused action is explicitly allowed to end a run early with declare_task_succeeded, it can now include skip_post_run_checks=True for cases where the task was already complete before the normal path produced attachments or output data.

Cyberdesk skips executing the configured Post-run Checks in that special path, records each check as successful with a message explaining why it was skipped, and does not generate a reusable trajectory for the shortcut branch.

What changed

GPT 5.5 is now available in Cyberdesk model configuration and is the new default model for the main workflow agent.

We have found GPT 5.5 to be a useful combination of smart and fast for computer-use workflows. It is especially strong when teams need responsive desktop actions without giving up too much reasoning quality.

Where Gemini still leads

Model choice is still task-dependent. For GUI grounding, Gemini 3.1 Pro and Gemini 3 Flash continue to take the lead in our testing. They remain strong choices for workflows where visual localization and screen grounding matter more than raw end-to-end speed.

Cyberdesk model configuration lets teams use this mix intentionally: keep GPT 5.5 as the fast default, then override specific focused actions or extraction steps when another model is a better fit.

What changed

Post-run Checks add workflow-level verification that runs after the main desktop automation steps complete. They are designed for cases where clicking through the UI is not enough and the workflow needs to prove that the expected result exists.

When a run has eligible checks, it can enter Running Checks after the main execution finishes. The final terminal status is decided only after those checks complete.

Initial check types

The first set of checks covers concrete artifacts and structured results.

• Attachment Exists verifies expected exported files, PDFs, CSVs, generated documents, or saved screenshots.
• Image Check uses a model to inspect screenshot attachments against a narrow visual rule.
• Output Data Passes Schema Validation is automatically managed for workflows with output schemas.
• Output Data Check evaluates structured output against a semantic rule when schema shape alone is not enough.

Why it matters

In production workflows, completion and correctness are different things. A report workflow may finish without exporting the PDF. A form workflow may end on the wrong confirmation page. A data extraction workflow may return JSON that has the right shape but is missing the important meaning.

Post-run Checks close that gap by making verification part of the run lifecycle and by storing check results on the run for review, webhooks, and downstream automation.

What changed

Cyberagent gives teams a faster way to move from "this is how the work is done" to a usable Cyberdesk workflow. Instead of starting every workflow from a blank prompt, teams can use examples, runs, and trajectories as source material.

The goal is to make workflow authoring feel closer to pairing with an operations engineer who understands desktop automation patterns.

What it can do

• Create workflow drafts from a video walkthrough of a desktop process.
• Edit existing workflows by translating requested behavior changes into clearer instructions.
• Analyze completed runs to suggest prompt improvements, missing checks, better outputs, and cleaner recovery paths.
• Review trajectories for replay candidates, brittle steps, unnecessary actions, and safety issues before approval.

Why it matters

Desktop workflows improve through iteration. A first draft gets the basic path working. Real runs expose ambiguous screens, missing outputs, unnecessary clicks, and places where the agent should stop for review.

Cyberagent shortens that loop by helping teams turn operational evidence into concrete workflow changes.

What changed

We added Get Main Instructions, a focused action tool that lets an agent reread the main workflow instructions before recovering and continuing a task.

This makes it easier for focused agents to handle unexpected screens or branches without losing the larger workflow context.