Thanks, those don't deterministically prevent the main loop from using tools thought, unless I'm wrong that's just prompting the main agent on when to use specialized sub agents
Thank you for your feedback - I have it working on iPadOS 26.5 (Chrome and Safari). The browser auto-play policy does require some interaction with the page before it will play (so the first key press is often mute, but should sound after). There is a button on the control bar to enable sound explicitly. I will look for more instances (not to ask the really obvious question, but do you have sound turned up or a bluetooth connection you're not aware of by any chance?). Thanks again
For https://www.asmusictheory.com, I built in a spaced repetition test section to aid memory retention, along with "a free play" mode for when you just want to explore.
Awesome, this is a great resource. I've been working on https://www.asmusictheory.com , which focuses excercises around an interactive onscreen / midi capable keyboard (where it makes sense). I'm building it as I learn myself.
Prior to this release, Agent Kanban - the online task board with agent harness integration for task management and context capture worked primarily with Github Copilot. With this release, the extension now supports Claude Code (the VS Code extension or TUI)
I also built a kanban board with agent integration and context management, with a vs code extension to go with it (also helps with git worktrees too): https://www.agentkanban.io
There is AS Notes - an Obsidian / Logseq / Roam alternative for use a s a VS code extension (is designed for use behind corporate firewalls, git friendly): https://www.asnotes.io
AI has been a great 'exoskeleton' for me. I fortunately had some good infrastructure and solid application base templates from before AI 'got gud' and so building on these has been the best of both worlds - a solid base and improved speed of development.
The UK is still the 5th biggest economy in the world. Public infrastructure feels like it's under huge strain however, and there is also a big problem with inequality, which seems to be changing under Labour, albeit slowly.
AI is beat thought of as an exoskeleton, you'll be at a huge advantage if you learn how to use it properly, and you will, unfortunately fall behind if you don't. I still think we're going to need people who can reason about code, and the amount of code to reason about is exploding in volume. Think of it as doctors having access to better drugs and techniques - they can can cure more illness, but the bar and expectation of what they can do will just raise. And doctors are still well paid, because what they do is important and needs doing well.
I've wondered this too - exactly how are our inputs and outputs useful as training data? So I asked Gemini. Apparently using negative sentiment in user or llm responses can serve as RLHF, and the human prompts can also serve as useful data for what problems the llms need to be able to solve. There's also that smaller models can train on and improve from data from larger models but that's less relevant when not switching models in context.
> A human iteratively (sometimes painstakingly) shapes and reshapes their creation until it sufficiently matches what’s in their mind’s eye. The odd thing about generative AI is that it can produce substantial form with minimally applied intent.
To add - I think attention is the scarce resource on both ends. Was this creation important enough to give it my full attention (i.e. do it my self). Is this creation important enough for the user to use / consume (i.e without an agent interacting with it on our behalf or summarising)
Same, with regard to TUIs in general. The VS code copilot chat extension has really nice integration for 'human in the loop' style agentic development. I build some tooling - https://www.agentkanban.io to integrate a taskboard and git worktrees with copilot chat
There's a few apps out there that facilitate handing off to agents from kanban boards. I needed something more 'human in the loop', handing off to an agent without good visibility of the change set and opportunity to steer doesn't work for me. https://www.agentkanban.io links a taskboard with github copilot chat in vs code via our extension so we have the benefit of task management and context capture from the chat to the tasks. This gives us all the features of a top harness (vs code) and the task / project management features at the same time.
I tried antigravity when it was first released, I didn't see an advantage over vscode, which it's forked from, and there were a few extensions I used that aren't supported. I've been a huge fan of copilot in vscode, the tight integration beats the TUI harnesses, and I've built some tooling around it (https://www.agentkanban.io) so I've got an integrated task board, context capture on the board and git worktree management for parallel tasks)