magospietato

Haven't built anything on the agent skills platform yet, but it's pretty cool imo.

On Android the sandbox loads an index.html into a WebView, with standardized string I/O to the harness via some window properties. You can even return a rendered HTML page.

Definitely hacked together, but feels like an indication of what an edge compute agentic sandbox might look like in future.

It's not that the context window is adequate, but rather an agentic LLM can search the source of truth using appropriate tools (SQL, term search, etc.)

RAG made sense when the semantic search was based on human input and happening as a workflow step before populating context. Now it happens inside the agentic loop and the LLM already implicitly has the semantics of the user input.

Living here the decline is tangible. And this is West Oxfordshire; not one of the poorer parts of the country.

An example in microcosm: a local village suffered road flooding due to failed maintenance of water pipes. Our rent-seeking privatized water company effected the minimum repair required by regulation.

The next section of old pipe burst almost immediately, flooding the road further for most of January, utterly destroying the surface, through the road base in many places. Even at a crawl it's difficult to avoid tyre damage.

Over a month later the water repairs were effected. Then shortly after some local roadwork notification signs were put up.

Those expecting repairs to the moonscaped road were disappointed: instead the relentless bureaucracy of British local government installed traffic calming measures on top of the broken road, as the work had already been booked and could not be stopped by any means as even basic roadworks lack any degree of dynamism in their execution.

All this still needs to be made right. These small scale failures will compound and compound until the entire state is drowned in the consequence of its incompetence.

Context caching deals with a lot of the cost argument here.

Con: from 2016 w. no followup?

Surely the skill for a cli tool is a couple of lines describing common usage, and a description of the help system?

Expires after 60 days per the link. In case you're thinking of stretching it out beyond a couple of months.

Complete aside, but it's beyond infuriating I need to enable a VPN here in the UK to view this link.

Having seen a family member die absolutely horrifically in a matter of weeks due to late-diagnosed pancreatic cancer, I'd consider suicide if I got the same diagnosis.

I wonder what training insights could be gained by having proven general intelligences actively navigate a generative world model?

Always wondered if dreaming is some kind of daily memory consolidation function. Logged short-term/episodic memory being filtered and the important bits baked by replaying in a limited simulacrum.

I still manually approve tool use requests at the start of a run. As it gets deeper in I might allow it to run safer commands without that oversight (e.g. writing to local text files), but potentially destructive execution still requires approval.

As for the local env, I'm treating the Android terminal as a sandbox. Anything gets trashed I just reset and reinstall my toolchain.

I won't pretend I'd use this workflow for anything high-stakes. But for simple things like "I wonder how my Hue lights actually work?", its viable.

Possibly Codex, but I've only used Claude Code so far.

Worth pointing out I'm not SSHing to a different device. Claude Code installed and running directly in Android terminal on my phone.

I've built ASP.NET Core APIs on-device this way. Install dotnet in the terminal and Claude can write code, build, run unit tests, and even run the API on localhost. Then use `git` and `gh` to commit, push and raise a PR.

This is how I do mobile device coding. Android terminal w. git and gh installed and authenticated. Claude manages the feature branching and PR process; I review the PR in the GitHub mobile app.