I think standard software development practices are your friend here. Separate project spaces with atomic tickets. Separate documentation for planning /reference. Clear guidance in the tickets for appropriate context.
Separate delegated sessions for projects against those tickets.
@mattpocockuk Spawning a session that's running in the background is handy. Naming sessions is a valuable practice any time. By giving it a name, if you want to inspect it later or jump into interaction, you will be able to find it easy by name when running --resume.
@zarazhangrui Absolutely! AI can run your whole content pipeline. I vibe code entire articles, drawings, graphs, editorial images, LinkedIn summaries, LinkedIn slides. A lot of it from html. Can send that same html to hyperframes for video. All of it with AI workflows.
@BrianRoemmele Don't you think it will more likely be picked up from an inference provider like Open Router?
Local seems unlikely to me as there is capex and operation maintenance overhead. Cloud is well positioned to deliver competitive reliable service at scale.
This looks promising as a remote control app. I'll get it installed and test drive it but would be especially impressive if it has multi-session per agent.
Would really like a local agent as well tethered to the remote so they could coordinate and the local phone agent could access local phone data.
Sometimes the llm picks the take action instruction instead of following the more complex loop that has delegation and supervision.
You ask your model what instructions it followed to take an action. This will help reveal conflicts in harness files.
When a agent seems to ignore you it's likely there are conflicting messages within the harness. Not because the platform nurfs the service
I have a very good loop skill, but sometimes my agent skips it because it also has a "if you can take action, do it" instruction.
"You are right. I didn't even read the skill file and sort of went from vague memory even though you explicitly told me to exactly follow the skills in full caps"
New model or Fable 5 back imminently. There is no inference available for existing stuff. Godammit Anthropic need more compute. It's farcical
@petergyang@abustamante@OpenAI Everyone already citing many Enterprise businesses where product is the cloud, the through put, the support, etc. Doesn't seem a big deal and more likely every day.
@BrianRoemmele Was thinking of you when I found this house. There was a liveliness that was expressed in everything mid-century!
https://t.co/URoywBYQ3w
🧠Write more about whatever you think on X
Note to myself and follower friends
I've seen multiple cases of very big accounts launching some trend when others were already writing about it and doing it months earlier. At least post for a track record of being early.