Today's GitHub Trending highlights a clear shift: agents are moving from "can chat" to "can operate."
-
trycua/cua— Computer-Use 2.0 with open-source drivers and cross-OS fleet support -
BuilderIO/agent-native— A framework for building agentic apps -
coder/coder— Secure environments for developers and their agents
These are exciting projects. But we learned a painful lesson while building our own automation pipelines: "can operate a single UI" is not the same as "can reliably run an unattended business workflow end-to-end."
The gap we hit
Here's a scenario we ran into: we wanted to automate a daily data collection workflow — open a browser, log in to an internal portal, navigate to a report page, download the file, parse it, and send a summary to a Slack channel.
Individual steps? Agent handles them fine. But string them together into an unattended workflow, and things break:
- The portal throws a session timeout mid-flow — no retry logic
- A file download dialog changes — the Agent doesn't know what to do
- The workflow fails at step 7 — no checkpoint to resume from
- You want to run this on a schedule, headless — no built-in scheduling
The missing layer wasn't "can the Agent click buttons." It was: flow orchestration, exception handling, retry, and state management.
What we needed: an Agent-ready RPA layer
This is exactly what iflytek/astron-rpa is designed for:
- Out-of-the-box automation tools — desktop and browser operations as composable steps
- Agent-ready architecture — Agent handles decision-making; RPA handles execution
- Built for unattended execution — exception handling, retry, and state checkpoints are part of the workflow, not afterthoughts
- Individual & enterprise ready — from personal automation to team-scale deployment
The full stack: orchestration + execution
For more complex scenarios, pair it with iflytek/astron-agent:
| Layer | Repo | Role |
|---|---|---|
| Orchestration | astron-agent |
Decompose complex tasks, manage multi-agent workflows, handle state |
| Execution | astron-rpa |
Execute desktop/browser operations, handle UI interactions, retry on failure |
Think of it as: astron-agent decides what to do; astron-rpa actually does it.
The takeaway
Computer-Use 2.0 projects like cua are solving the infrastructure layer — and that's great for the ecosystem. But if you're trying to automate a real business process that needs to run unattended, you need more than an Agent that can click buttons. You need a workflow execution layer with proper error handling, retry, and state management.
That's the gap astron-rpa fills.
Links:
-
iflytek/astron-rpa: https://github.com/iflytek/astron-rpa -
iflytek/astron-agent: https://github.com/iflytek/astron-agent
Top comments (0)