@burkeholland Less verifiable for sure. Backend changes that pass (good) CI gates are pretty much guaranteed to not be catastrophic. Frontend changes can lint, typecheck, pass tests, pass automated a11y, whatever, and still be a train wreck.
@ClintRutkas Half the time I have fundamental
disagreements with what they consider to be correct since itβs some arbitrary idea of whatβs βrightβ.
At worst, this saved me an hour or two of work figuring out when to sell stuff. At best it saved new a few thousand dollars of getting it wrong by hand. Nice.
Them: Check out this skill I checked in
Me seeing a 700 line markdown file that could be a script: ok bet.
Expensive? Yes.
Paying this once to not run 3x the inference on every PR? Worth it.
@burkeholland@LukeParkerDev :) selfishly, I'm driven mad when there's a group chat that has all the answers and I didn't even know it existed.
I'll more than concede that the Teams channels UX is *far* from good which leads to a lot of your valid frustrations.
@burkeholland@LukeParkerDev Nobody you donβt invite can know that group chat exists or find any of the information in it. Itβs a risk and terrible for agents to learn and gather context that way.
Similar point to what Shopify did forcing their coding agent to be in public slack channels.
@LukeParkerDev@burkeholland Except everyone is allergic to using Channels so everything is buried in random group chats that are the devil's spawn where information goes to die.
Slack being opinionated about channels is so important to it being good at scale.
@burkeholland Take is 50% vibes, 50% reports like this consistently showing cost per task is significantly lower for 5.5 than Opus, despite having effectively the same per token cost. https://t.co/IjbdnrXyD4
We benchmarked the GitHub Copilot agentic harness against the harnesses that ship leading models natively.
Holding the model and task fixed across SWE-bench Verified, SWE-bench Pro, SkillsBench, TerminalBench, and Win-Hill, the results were clear:
β Task resolution on par with model-vendor harnesses
β Fewer tokens across most configurations
π‘ A key learning: With GitHub Copilot supporting more than 20 models, you're free to pick efficiency or peak quality per task.
@jeffwilcox@GHchangelog Glad this is one of the few things I can rely on being nearly instantaneous enabled at the same time as the rest of the world. Unlike most things at The Company π