dmj

https://www.lg.com/us/terms i.e. use these in schools and the students can't sue LG, only the school.

Agent skills are basically markdown text files that you can write, or more likely get an agent to write, that have a bunch of rules eg don't do X on production, always ask about Y. You can have the coding agent run them when they run. Not infallible but good thing for engineers https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf

Already happening. I had an 'incident' a few months ago where Claude was trying to get some terraform working building some infrastructure, figured it would be simpler to delete the GCP projects and re-create them (should have been fine, nothing is designed as a single point of failure especially infrastructure). Off it went, until it stopped because it deleted the project it was running in. We've added some guard rails to our agent skills since then.

yeah for repeatable things Configuration as Code is the way forward - Automation tools such as Ansible/Puppet are good here too, especially with an AI writing that config.

I've been using Claude Code, Opencode and Pi. I'm starting to gravitate towards pi over the last couple of weeks. Although these are 'coding harnesses' they typically use a lot of system commands to understand and manipulate their environment - combined with a half decent model they are well suited to system tasks, and I am sure there are a number of OS specific skills around to give them a hand.

This really should be with your legal team. Contract law is notoriously difficult to understand and they will (should!) have a good understanding of how to pursue this.

Perhaps llama RPC and share the model across all the hosts it's experimental but worth a shot https://github.com/ggml-org/llama.cpp/blob/master/tools%2Frpc%2FREADME.md

I have done this, but not recently. I don't recall there being any issues at all TBH - could it be something simple like a cable failure ? Can you boot it via USB?

answered my own question: https://www.kenable.co.uk/idc-punch-down-to-rj45-plug-for-cat6-solid-ethernet-cable-connector-4-pack-012938?srsltid=AfmBOop7f4JCooG0fCN3ZTph8CLhll0xaAHq9khBOcApdzdD6efjIT7ORRo

Doesn't using the RJ45 coupler mean you then need much more expensive stranded cable in the walls? Thought the whole idea of punchdown was for single core and that had better signal/noise over stranded - or can you get male RJ45 for single core now ?

@dhicks Apparently The Alibaba qwen 3.8 27B model is very good for local work. I tried it on a 24GB M4 pro (ollama and opencode) and instantly ran out of memory, but it might work on a 48GB ? Here's a good writeup (on a 128GB M5) https://simonwillison.net/2026/Aug/16/qwen-38-27b/

It could cut either way I think: Teachers rapidly creating applications and selling them on the open market without any real understanding of software development will end up with the same poor quality no matter how good the agent is if the infra, QA, secops etc isn't behind it. I dread a vibe coded SIMS replacement. Or more optimistically: some high quality OSS projects that leverage people who do have those experiences and governance to produce something good for schools to use. Supported by companies that spring up around that software. Sadly I expect the former.

That's how we did it (10yrs ago). It's not difficult to setup. At it's simplest just route all the traffic through the school.

A good analogy is spreadsheets. In the olden days departments would build their own dashboards, perhaps with a Javascript/VBScript. The spreadsheet application, along with the IT team, would take care of all the deployment/security etc. So IMO its as much a question about how to handle the ops side of it (the apps that serve / deploy the AI apps) as it is the apps themselves.

We were doing Chromebook 1:1 back in 2014-15. It was successful in the school I was at (actually reduced support times as we minimised our windows estate). Last I heard it was still running. The reason it's not a national strategy is simply politics

We are struggling with the same issues, and have a number of these applications deployed (through Google Cloudrun). The fundamental problem is that people without development experience vastly underestimate how time consuming it is to maintain their applications. At the start the applications start as a 'nice to have' and that the owner will look after. Slowly they become business critical, and parts of the business genuinely cannot function without them. At this point you need to add in SLO's and metrics to make sure the app is behaving and has a good uptime - but the 'developer' didn't understand this requirement. Then a library that the application is using has a critical vulnerability, but the 'developer' didn't add any build pipelines so there's no way of knowing, because the school doesn't have the infrastructure in place to manage automated builds and scanning. Once the library has been updated, the application starts behaving weirdly, or stops working entirely- the schools didn't have any regression testing in place and the 'developer' didn't think to add it or maintain it. Then the 'developer' leaves and the rest of the IT department are lost. IMO don't block this - instead insist on: 1) using a git repository 2) Docker builds 3) automated builds with a) linting b) container scanning 4) automated deployments 5) dependabot/renovatebot to automate code updates 6) unit testing 7) regression testing before deployment Once the PA has figured all this out, deploy the application then ask why s/he's not earning £80k as a devops engineer.

With regards to creating git repos, cookie cutter might be a good option for students to clone, with the basics in there https://github.com/cookiecutter/cookiecutter You can drive the cli from your agent and it will use fewer tokens

I'd love to do that, but I'm pretty sure to get anywhere close to gemini/Claude/GPT you'll need a £250k rig. Closest next best thing is GCP hosted but at $148 per hour ...

A good method is billing via GCP enterprise agent platform (formerly vertex). You can setup a GCP project and enable the necessary API's, then Claude Code/Opencode can bill directly through the GCP project. (not sure about Antigravity, but given it is Google I would be surprised if it didn't) There are some models as a service and pay per token (as opposed to hosting your own model). All the Gemini and Anthropic models allow this along with some of the openweight models. Openrouter is also a popular choice.

AFAIK there's not an equivalent to https://code.claude.com/docs/en/auto-mode-config in antigravity, last time I looked it was 'dangerously skip permissions' or click each time. Opencode is a lot more flexible FWIW.

As you are a 'Google School' IIRC. I would suggest looking at CloudRun fronted by Identity Aware Proxy - that how we are doing that. WRT the vibe coding aspect. I'm currently flitting between Zed (editor) using GPT-5.6-sol and Claude Code (opus 5). Occasionally I use OpenCode with either of those models. Not looked at antigravity for a while, when I did I (personally) found the Gemini models the worst out of GPT/Sonnet/Opus. I'm quite keen to get my hands on Kimi3 and GLM-5.2 my company is so far restricting us to the U.S. models, which is weird because they are more restrictive and more expensive.

coyote works with ACME 😁 https://github.com/blendbyte/coyotecert

If it's causing this much hassle, you can buy a generic 1GB up/down connection on a business rate for £50/month. For the cost of a laptop I'd just suck up the cost, wait for the least line to expire then get a contract elsewhere.

https://www.legislation.gov.uk/ukpga/2026/21/section/36/enacted

Serverless simply means the servers are ephemeral (typically containers that scale from zero at night to full load in the day). So as long as the data meets the 3-2-1 rule it is compliant, the servers themselves don't count because they get built and destroyed multiple times per day. In this scenario onsite would be in the same provider as your cloud infrastructure, and offsite would be a third party cloud. Ideally you would also have an Isolated Recovery Environment (IRE), i.e the third party cloud also spins up the same containers with the same data - so if the entire infrastructure can be replicated incase of a provider failure.