LunarWing 2 Release and a new development paradigm
LunarWing 2
Why is LunarWing 2 so exciting?
LunarWing 2 is many things, but one of the most exciting things to me isn’t the new features.
It’s not the new agentic loop engine with LLM streaming and interruption support.
It’s not the revamped MCP system.
It’s not the self-improving agent skills, the new onboarding UI application, or the revamped, redesigned snappier UI.
The most exciting thing, to me personally, is the way we’re now building LunarWing and a demonstration of agent capability through sheer experimentation…
Autonomous development loop
One of the new things we’ve been playing with is how we’ve built a specialized autonomous development loop that runs as a routine.
We give an agent a list of 10-20 very specific, clearly defined goals, and each is parsed by a subagent.
The agent operates and controls a coding harness (which may use a very specific configuration seperate from that of the agent itself) in order to work on a single goal. The end result is pushed to a git branch, a PR is opened to keep track, a checklist item is checked off, and the script exits.
The work is logged and saved. Failures and metrics are recorded and constantly evaluated. The metrics system built for this is impressive. This entire single routine will warrant an entire blog post unto itself in the future.
Automated integration testing using multi-tenant environments
Similar to the autonomous development loop, we’ve been able to build a similar system in order to do automated testing. This is essentially a CI/CD platform on steroids with agents utilizing automated browser controls in order to test features in a specific set of branches.
It has automated cleanups and is able to test multiple instances at a time. It’s a game changer for the way we test our code.
Other thoughts
We’ve been using these new, experimental routines since prior to 1.1.9.0 and we continue to experiment and improve them. Our routines for these have been refined and customized to the development of LunarWing itself, but we believe that in the future they could be adapted for agents to work on other kinds of projects.
We’ve also been working on more advanced memory routines based on research and development that has been ongoing for several months. Utilizing simple self-correcting memory routines, we’ve been able to reduce hallucinations by over 80%. This kind of difference is noticable when LunarWing is compared to other agents. We hope to include these into the core project at some point as well.
Release
Today, we’re releasing v2.0.0.0 - the first release of version 2.
We’re no longer using Github for releases or accepting contributions and we’ve transitioned to a new repository
Release Notes
As usual, the release notes are long and detailed because we value transparency and our roadmap for future releases has been updated.
LunarWing’s release notes don’t look like that of most other open-source agent projects because we take pride in our work.
You can find these on the release page as usual.