So, I have been working on my project called isol8, which is a power packed, fast, safe and versatile sandbox environment for ai agents to execute code.
--
Now, to deliver all the features quickly, I have been using various ai coding agents to deliver a whole lot of features and squashing bugs. I still have an open list of things I need to work on (github issues).
To do this, I have now settled on an approach which works for me, and I am going to share them today. I’ll also publish more parts to this as I figure out more things, but here’s were I am right now.
Tools
OpenCode- I love this agent harness, the output has been really good, and I can choose what model I want to use. Highly recommend.Github Copilot- Not the harness but just the model inference. I have tried multiple llm providers - openrouter, opencode zen. But Github Copilot provides enough usage in their pro+ plan (see all plans). 1500 monthly calls to Opus, sounds really good to me ngl and OpenCode supports it natively 🤌.
Workflow
- I heavily use github issues to store ideas and bugs. I would just discuss things with as many people and llms on what could be the possible issues regarding security or performance, and then ask it to follow proper templates to create issues on github using gh cli.
- I also have proper guardrails, I am using ultracite for formatting and linting, conventional commits and type checks as pre-commit hooks. Even for github pull requests, bug reports, feature requests I have setup proper formatting.
- I have setup a few skills -
- docs-quality - to write the comprehensive documentation, looking at existing code, existing documentation and figuring out the best way to display it. It maintains parity with code features.
- issue-to-implement - best of all, this one just takes in an issue url and starts the analysis, figures out a plan, responds back on the issue thread, fix the issue / implements the feature, adds more tests to cover everything, updates the documentation and then finally commits everything and raised the Pr on github with the right description format and labels.
- project-issues-batch - takes in github projects url and uses gh cli to figure out associated issues. then decides what is dependant on what and then invokes sub-agents to implement everything and raise PRs per issue.
- finally, we have test coverage. I have intensive unit and integration test coverage, so that no buggy release goes live. and we also have regression test suite which runs after every release.