DEV Community

Discussion on: Ran out of Cursor tokens and switched to GitHub Copilot: Side-by-Side

Collapse
 
maxxmini profile image
MaxxMini

Interesting comparison! I hit the same token limit wall with Cursor and ended up going a completely different direction - CLI-based agents (Claude Code) orchestrating sub-agents for parallel tasks instead of relying on IDE integrations. The biggest win was automating the workflow around coding: draft generation, deployment verification, health checks, all chained together. Went from manually doing everything to 80+ automation scripts in a week. Curious if you have tried any non-IDE approaches? The Plan Mode frustration you described is exactly why I moved away from IDE-embedded AI - too many layers between intent and execution.

Collapse
 
maximsaplin profile image
Maxim Saplin

I use OpenCode quite often, even created a skill to launch resumae subagents powered by OpenCode. CLI tools are great if you want to build your own pipelines for sure. In terms of general use, see no major difference in capability, be it GUI or TUI. Question of preference and token availability

Collapse
 
maxxmini profile image
MaxxMini

Good point about GUI vs TUI being mostly preference — I agree there's no fundamental capability gap anymore. OpenCode looks interesting, hadn't seen the resumable subagent pattern before. How do you handle context window limits when your subagents are working on larger codebases? That's been my main pain point with CLI agents — they burn through tokens fast on big repos even with good .clinerules or AGENTS.md scoping.

Thread Thread
 
maximsaplin profile image
Maxim Saplin

Frankly I have not seen subagents to ever overflow context, that's not that easy, did multiple attempts and still exploring long-horizon subagents orchestration: github.com/maxim-saplin/hyperlink_...

So far implementations I saw (runSubagent tool in GH and task in OpenCode) do simple request/response disposable dialogs where an agent sessions is started and than the main agent spinning the subagent only cares about the final result.