📣We are pleased to announce that @ask_ui is participating as an Exhibition Sponsor at the 3rd European Automotive Intelligent Cockpit and Smart Driving Summit 2026.
More info:
https://t.co/JdGojzjy7G
Email: [email protected]#IntelligentCockpit#SmartDriving
We just shipped AskUI Desktop
"start navigation to 1600 Amphitheatre Pkwy"
"expect Google Maps shows an active route"
that's the entire test. the agent sees the screen and does the rest. 🧵
and it's not just pass/fail.
every run comes with a full, audit-ready report.
Every step, every screenshot, every decision traced.
You see exactly what the agent did and why.
why this matters:
rule-based automation only handles what you planned for. one UI change and your test suite breaks.
an agent that actually sees the screen finds the button wherever it is
It's not a script runner.
It uses an LLM to read the intent of your test, decide which tools to use, and execute.
Screenshot → API call → Shell command → MCP server
And Selenium only covers the browser. APIs.
Shell commands. Car dashboards. HMI panels.
The moment your product goes beyond the browser you're adding another tool.
And another.
Last batch and probably the biggest for flexibility.
Agents can now read files and list directories on the automation target directly.
And AskUI now supports any OpenAI-compatible model. Ollama, vLLM, LM Studio, Together AI, RunPod. If it exposes a chat completions endpoint, it works.
Bring your own model.
pip install askui --upgrade
https://t.co/to4g0vXHPZ
#agentdev
Two things we'd been putting off.
WebAgent is now a proper browser automation agent full Playwright tools, not just navigation. Mouse, keyboard, screenshots, the works. WebVisionAgent is deprecated.
And AutomationError: a new exception type for genuinely unfixable errors. Missing credentials, unreachable services. Stops execution immediately instead of sending the agent into a retry loop.
AI & Agentic Workflow News: March 18–25
Most are fighting over model IQ. The real bottleneck is the execution layer.
This week marks the end of the Mac Mini era for agents. Between Anthropic's native remote execution and GitHub's 12-month LTS, we are moving from how to host to how to execute anywhere.
But as builders know, a brain is useless if it can't bypass a lock screen or survive a session drop.
Full breakdown on the shift to Unstoppable Execution:
#AIAgents #Infrastructure #Automation #SoftwareEngineering
The model is the brain. The execution layer is the nervous system.
3 releases, one thesis:
- v0.25: Architecture rewrite speaker-conversation pattern with pluggable cache execution
- v0.26: One agent controls PC + Android simultaneously
- v0.27: Execution cost tracking + OpenTelemetry tracing (shipped today)
The full breakdown ↓
#Agents #AI #ComputerUseAgent #AskUI