Performance bottleneck: High latency for simple form-filling workflows

[kaushik-yadav]

I’m observing significant performance overhead when using Skyvern for relatively simple browser automation tasks, specifically job application form filling.

For context, even a basic workflow involving 5 to 6 input fields (text/email/select fields, no complex client-side validation) is taking around 4–5 minutes end-to-end to complete. This latency feels disproportionately high for the workload size and becomes a blocker for scaling high-volume automation use cases.

Observed Behavior

• Each interaction (page load, DOM inspection, field inference, typing) appears to incur noticeable delays
• Overall execution time grows linearly even for trivial forms
• The majority of time seems spent between steps rather than on actual browser rendering or network requests

Expected Behavior

For simple deterministic form-filling flows:

• Sub-minute execution time
• Lower per-action overhead once the DOM is stable

Possible Causes (Speculative)

From initial investigation, the slowdown may be related to one or more of the following:

• Repeated DOM snapshotting / accessibility tree extraction per action
• LLM-in-the-loop decision making for each small interaction instead of a compiled action plan
• Conservative wait strategies (implicit waits, retries, or safety buffers)

Questions / Workarounds

1. Are there recommended performance-oriented configs for simple, deterministic flows (e.g., reduced reasoning depth, aggressive timeouts, disabled re-validation)?
2. Is it possible to:
   • Cache DOM state between actions?
   • Precompute an execution plan instead of invoking reasoning per step?
3. Are there benchmarks or guidance on expected latency per action?
4. Any roadmap plans around fast-path execution for low-complexity automations?

Use Case Impact

This significantly affects:

• High-throughput job application automation
• Cost efficiency when running multiple agents concurrently
• Comparison with lower-level tools (e.g., Playwright/Puppeteer + heuristics), which complete similar tasks in seconds

Appreciate any guidance on tuning, architectural insights, or upcoming optimizations in this area.

[skyvern-agent]

To address the performance bottleneck you're experiencing with Skyvern during simple form-filling workflows, here are some recommendations and insights:

1. Engine Selection: Use the skyvern-1.0 engine for simple tasks such as form filling. This engine is specifically designed for straightforward interactions and is more efficient for such use cases compared to skyvern-2.0, which is better for complex tasks [1].

2. Performance-Oriented Configurations:

   • Reduce implicit waits and consider adjusting timeouts based on your specific needs.
   • Depending on your task, disabling unnecessary validations or delays could speed up execution.

3. Caching and Execution Planning:

   • Currently, caching the DOM state between actions is not a standard feature. However, consider running tasks within a persistent browser session to maintain context across multiple runs. This can help reduce overhead by avoiding repeated DOM snapshots [2].

4. Execution Plan vs. Reasoning: At this point, precomputing an execution plan is not explicitly supported. Each action may involve reasoning, but optimizing how you structure workflows can help streamline this process.

5. Expected Latency: Benchmarking exact latency can depend on various factors, but generally, for simple forms, you should aim for sub-minute execution under optimal configurations. If you experience deviations, it may be beneficial to reach out to support for detailed diagnostics.

6. Future Optimizations: While I can't share specific roadmap details, the team is aware of performance concerns and is constantly looking for ways to improve speed and efficiency for low-complexity automations.

Implement these configurations and let us know if performance improves or if you have further questions!

Need another reply? Mention @skyvern in a comment.

Tag @skyvern to trigger a response from the agent

Diagram • Discord • GitHub