{
  "$type": "site.standard.document",
  "content": "---\ntitle: \"The road to COMP4020: providing sharp tools\"\ndescription: \"Getting 200 students adequate access to frontier coding agents is the hard\n  logistics problem behind COMP4020.\"\ntags: [comp4020]\n---\n\n:::tip\n\nThis post is part of a series I'm writing as I develop\n[COMP4020: Agentic Coding Studio](/blog/2025/12/19/comp4020-rapid-prototyping-for-the-web/).\nSee [all posts in the series](/blog/tag/comp4020/).\n\n:::\n\n:::info\n\n**Update:** this one worked out---Anthropic came through with $500k in API\ncredits for the class. See\n[Anthropic comes to the party](/blog/2026/04/02/anthropic-comes-to-the-party/).\n\n:::\n\nI was at the Australasian Computing Education Conference (ACE) in Melbourne last\nweek presenting my work on [LLMs Unplugged](https://www.llmsunplugged.org) and\n(unsurprisingly) a lot of the discussion was around the use of LLMs in the\nclassroom.[^llms]\n\n[^llms]:\n    There was a running joke where the few presenters whose papers _weren't_\n    about LLMs made a point of that fact in their intro, and the rest of the\n    presenters sortof gave an apology that theirs was yet another paper on LLMs.\n    But I digress.\n\nOne fun anecdote from\n[Claudia Szabo](https://www.adelaide.edu.au/directory/claudia.szabo) was about\nhow she gets her small kids to help in the kitchen, and that she gives them\nsharp knives and teaches them how to use them. And she wants the same for her\nstudents in the classroom.\n\nOne of my main concerns for teaching COMP4020 later this year (starting\nJuly 2026) is that I want my students to have a) access to the latest agentic\nLLMs and b) sufficient token budgets to use them properly. It's not fair to\nexpect students to shell out 200USD of their own money for a\nClaude/ChatGPT/Gemini Max plan; I have to provide them with a strong baseline\n(although my current plan is to allow students to use different models if they\nwish; it's too hard to police anyway).\n\nGitHub Copilot does have\n[free student accounts](https://docs.github.com/en/copilot/how-tos/manage-your-account/get-free-access-to-copilot-pro)\nbut they're limited to 300 \"requests\" (read prompts) per month, or even fewer\n(by a factor of almost 10!) if students choose Claude Opus 4.6. It's just not\nenough for me to run the sort of class and teach the sort of workflows I'm\nplanning to run.\n\nI'm not saying my students need\n[Gas Town](https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04)\nlevels of tokens, but they need to be able to regularly sit down and have long,\ninteractive back-and-forth sessions exploring different ideas and\nimplementations.\n\nSo, my current options are:\n\n1. get\n   [Anthropic](https://www.anthropic.com)/[OpenAI](https://openai.com)/[Google](https://ai.google),\n   or a company that resells those models like\n   [Microsoft](https://www.microsoft.com)/[GitHub](https://github.com)/[Amazon](https://aws.amazon.com)\n   to sponsor the class. I'm thinking that I'll need approximately 200 seats\n   (that's the number of students I'm expecting) and they'd need **at least**\n   the $20/mo pro-level plan, ideally the $100/mo level one---or similar amounts\n   of API credits\n\n2. get a partner with datacenter-grade hardware (e.g.\n   [Canva](https://www.canva.com), [LambdaLabs](https://lambda.ai), or ANU's\n   very own [NCI](https://nci.org.au)) to host an open weight model for us, e.g.\n   [Qwen3.5](https://huggingface.co/Qwen/Qwen3.5-397B-A17B),\n   [MiniMax M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5),\n   [Kimi K2](https://moonshotai.github.io/Kimi-K2/) or\n   [GLM-5](https://huggingface.co/zai-org/GLM-5) using [vLLM](https://vllm.ai)\n\nOption #1 is my slight preference because those folks are _really good_ at\nserving these models at scale. If we have to self-host then there's a risk that\nI'm on the hook if the model goes down one hour before the assessment deadline.\nBut it wouldn't be a disaster. The other bonus here is that it increasingly\nseems like the models are most effective with their \"default\" harness (e.g. Opus\n4.6 with Claude Code, GPT5.3 with Codex CLI, Gemini with the Gemini CLI). That's\nwhere new features land first, and that's where a lot of the post-training is\ndone.\n\nRe: option 1, Kathi Fisler and colleagues are teaching an\n[Agentic Coding course at Brown](https://cs.brown.edu/courses/csci1970kf/agentic-spr-2026/index.html)\nright now which uses a similar approach (even explicity using the studio/crit\napproach [like I am](/blog/2026/02/20/comp4020-the-core-mechanic/)). What\nthey're doing is re-imbursing students 20USD/month for a Claude Code Pro plan,\nand it\n[seems to be going well](https://bsky.app/profile/kfisler.bsky.social/post/3mfm5hpshds2j).\nMy preference is to not have to have students pay out-of-pocket and get\nre-imbursed; the ideal scenario is that I (as course convenor) can use the usual\nteam subscription management tools these platforms provide to give students\naccess when they enrol, revoke it if they drop or when the course is over, etc.\n\nIf you work at any of the above companies and would like to get in touch,\n[email me](mailto:ben.swift@anu.edu.au). I can offer the good vibes and\npublicity that comes with supporting the next generation of software developers\nand computer scientists in their learning. And I'm happy to share all the course\nmaterials online, inc. a shout-out to whichever model you end up providing.\nYou'll also have an opportunity to meet (if you like) the students, who are\nawesome and will be highly skilled and looking for work in the near future. And\nfinally, you'll have my gratitude :)\n",
  "createdAt": "2026-05-13T23:14:40.711Z",
  "description": "Getting 200 students adequate access to frontier coding agents is the hard logistics problem behind COMP4020.",
  "path": "/blog/2026/02/17/comp4020-sharp-tools",
  "publishedAt": "2026-02-17T00:00:00.000Z",
  "site": "at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.publication/self",
  "tags": [
    "comp4020"
  ],
  "textContent": "Getting 200 students adequate access to frontier coding agents is the hard logistics problem behind COMP4020.",
  "title": "The road to COMP4020: providing sharp tools"
}