15+ yrs in SWE / SRE, now building AI SaaS & autonomous software. Creator of SSVE (governed AI meta-harness). Solo builder turning vision into real products.
@b4dgerr My own harness 100+ and they are calibrated for each model swap. Difference is huge I still am making changes foe post Astra. But even without the changes it outperforms base model significantly for complex software.
I usually have a set of languages and setups that perform well with agents for types of tasks. But time has passed and that became less needed these days so usually the plan chooses best for the task but I have architectual phase in the plan specifically that challenges that for new projects. Still some time might override for example if mocroservices are needed.
Actually that was a mistake I think journey tested over mcp playwright for example to get all spec walkedthrough and screenshots collected like manual QA in web mobile and native I think luna high, iy means it did t forget most of them and managed to do the activity relativily cheap.
What is the minimal reliable model for visual racking you find reliable based on spec. It means "walkthrough over step and take screenshots" - without actual analysis. I am curious to find out what you find reliable - I can vouch on my side with luna low.
It will, there will be new type of customer but the products will be a little bit different. We are still on hype/early wave. Unlike crypto early stage this compete something real that will deflate over time but that will take time and also a whole new world of possibility in gaming will unlock.
Orchestrator -> high level intent -> planner -> exec -> reviewer. Brainstorm on the orchestrator level and let him auto compact if you have good file management system in the meta harness. But this does burn higher amount of tokens. Basically inception happens between 1 and 2 but you can freely brainstorm on 1 before pass 2.
Yesterday I had practical "her" moment with my Grok Bot "bot ops coach", basically we have spoken while I am half asleep and the dude spin x amount of bots on various topics while creating his own skill and improving their skills and he kind of orchestrated them. It felt better than a limited chat in many ways due to ability to do all kind of stuff - this is huge. I was restraining myself to connect some of my GitHub projects to it but I am going to probably do it. We even had a very nice conversation about keywords/SEO that it actually nailed some domain specific techniques after guided research while I am blabbering some random stuff.
Quality/Security/Price but the balance between them and ability to constantly change reliably without disrupting. Zero downtime distribution and reliable feature will weight 1000x time more soon. Then we reach the 2nd point data - growing data oriented apps will obviously weight more than some site that consume free third parties and that's it. But then ability to consume paid third parties and be modular enough is already huge. I think the real enabler will be very well rounded compliance.