{
  "$type": "site.standard.document",
  "content": "---\ntitle: \"The road to COMP4020: what's the theory here?\"\ndescription: \"There are lots of nascent theories for agentic coding---here's a survey, and\n  a plan to test them with 200 students.\"\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\nI was chatting with the CIO of a government agency last week and they asked\nabout this course. They're wrestling with this very issue: in the age of agentic\ncoding agents, how do I structure my teams, development workflows and QA\nprocesses?\n\nThe answer which I gave him, which I believe to be true in my bones, is that\nnobody really knows for sure. But a lot of folks are trying to figure it\nout---some in private conversations inside their organisations, some very\nloudly on (ugh!) LinkedIn, and lots in between[^loudly].\n\nThe question was really one of _theory_; while there are many in the hard\nsciences that would criticise software development/engineering as having a\nlamentably loose proof relationship between the theory (agile! scrum! 10x\ndevelopers!) and what works in practice, at least there are theories about what\nworks and what doesn't, and there's enough agreement about what these courses\nare for people to write textbooks and run university degrees. But since Claude\nCode was released in May 2025, you can feel the ground shifting.\n\nHere's my attempt at a survey. Some of these are genuinely impressive; others\nI'm still not sure about. I've split them into rough categories, though there's\nplenty of overlap---the best frameworks come with tools, and the best tools\nembody a theory about how work should flow.\n\n:::info\n\nNone of the resources below are \"getting started\" guides---if you haven't\nactually used any of these tools yet, each vendor has official onboarding docs:\n[Claude Code](https://code.claude.com/docs/en/quickstart),\n[Codex CLI](https://developers.openai.com/codex/quickstart),\n[Gemini CLI](https://geminicli.com/docs/get-started/), and\n[GitHub Copilot CLI](https://github.com/github/copilot-cli-for-beginners).\nThese are practical \"here's how to install and use our thing\" resources, not\ntheories about how agentic coding _should_ work---but they're worth running\nthrough before diving into the methodological debates below.\n\n:::\n\n### Frameworks and methodologies\n\nThese are the structured approaches---the ones with a name, a thesis, and\n(usually) a manifesto. They're trying to answer \"how _should_ we work with\nagents?\" rather than just \"how _do_ we work with agents?\"\n\n- John Regehr's\n  [_zero degree-of-freedom_ approach](https://john.regehr.org/writing/zero_dof_programming.html)---constrain\n  the agent so tightly that there's only one correct output\n- Simon Willison's\n  [Agentic Engineering](https://simonwillison.net/guides/agentic-engineering-patterns/)\n  guide---ongoing, \"kind of book-shaped\", and probably the most comprehensive\n  single resource right now\n- [spec-driven development](https://harper.blog/2025/04/17/an-llm-codegen-heros-journey/)---Harper\n  Reed's \"hero's journey\" post is the canonical intro, and the thesis is blunt:\n  \"the spec is the godhead\"\n- Jesse Vincent's [Superpowers](https://github.com/obra/superpowers)---enforces\n  mandatory design, planning, and TDD via composable markdown \"skills\" that you\n  bolt onto your agent\n- Anthropic's own practices, documented in their\n  [Claude Code best practices](https://code.claude.com/docs/en/best-practices)\n  guide and the\n  [how Anthropic teams use Claude Code](https://claude.com/blog/how-anthropic-teams-use-claude-code)\n  blog post\n- Birgitta Böckeler's\n  [harness engineering](https://martinfowler.com/articles/harness-engineering.html)---a\n  framework of \"guides\" (feedforward controls) and \"sensors\" (feedback controls)\n  for steering coding agents toward better output while reducing human\n  supervision\n- the Deer Valley retreat consensus, written up by\n  [Martin Fowler](https://martinfowler.com/bliki/FutureOfSoftwareDevelopment.html)---about\n  50 luminaries (Beck, Yegge, Gene Kim, etc.) locked in a room hashing out the\n  future of software development. Chad Fowler's framing: \"the rigour has to go\n  somewhere\"\n\n### Tools\n\nThings you can actually `pip install` or `npm install` and try right now. These\noverlap with the frameworks above---the best tools are opinionated about\nworkflow.\n\n- [Deciduous](https://notactuallytreyanastasio.github.io/deciduous/)---decision\n  trees for AI coding agents, so your agent's choices are queryable and\n  persistent\n- [Chainlink](https://github.com/dollspace-gay/chainlink)---a CLI issue tracker\n  designed specifically for AI agent workflows\n- [Gas Town](https://github.com/steveyegge/gastown)---Steve Yegge's multi-agent\n  orchestrator for running 20--30 Claude Code instances in parallel (named after\n  Mad Max, naturally)\n- GitHub's [spec-kit](https://github.com/github/spec-kit)---an official toolkit\n  for spec-driven development, agent-agnostic\n- [Plandex](https://github.com/plandex-ai/plandex)---plan-first CLI agent with\n  version-controlled plans and a sandbox for reviewing diffs before applying\n- Jeremy Howard's\n  [Solveit](https://www.answer.ai/posts/2025-10-01-solveit-full.html)---explicitly\n  \"the opposite of vibe coding\", all about small steps and deep understanding\n\n### Reflections from practitioners\n\nAnd then there's the \"I tried it and here's what I think\" genre. These are\nvaluable precisely because the authors have enough credibility and experience to\nsay something beyond \"wow, cool\"---though several of them do also say \"wow,\ncool\".\n\n- [Andrej Karpathy](https://x.com/karpathy/status/1886192184808149383) coined\n  \"vibe coding\" and later refined his thinking toward \"agentic engineering\"\n- [Ryan Dahl](https://x.com/rough__sea/status/2013280952370573666) (Node.js,\n  Deno) declared \"the era of humans writing code is over\"---2.3 million views\n  and counting\n- [Mitchell Hashimoto](https://mitchellh.com/writing/my-ai-adoption-journey)\n  documented the full arc from scepticism to productive use\n- [Charity Majors](https://charitydotwtf.substack.com/p/2025-was-for-ai-what-2010-was-for)\n  argues 2025 was for AI what 2010 was for cloud---but writing code was always\n  the easy part; observability and ops are where it gets real\n- [Armin Ronacher](https://lucumr.pocoo.org/2025/6/12/agentic-coding/)---practical\n  agentic coding recommendations from the Flask/Sentry creator\n- [Maggie Appleton](https://maggieappleton.com/gastown) on agent orchestration\n  patterns and why design and critical thinking are the new bottlenecks, not\n  code generation\n- [Cassidy Williams](https://cassidoo.co/post/vibe-coding-yawn/) found vibe\n  coding effective but joyless---\"there's no 'YAY I am a GENIUS' feeling\"\n- [Kent Beck](https://newsletter.pragmaticengineer.com/p/tdd-ai-agents-and-coding-with-kent)\n  on TDD as counterbalance to AI agents (\"agents keep trying to delete tests to\n  make them pass\")\n- [antirez](https://antirez.com/news/154) (Redis creator) on \"automatic\n  programming\" vs vibe coding---\"LLMs are good amplifiers and bad one-man-band\n  workers\"\n- [Jeremy Howard](https://www.answer.ai/posts/2025-10-01-solveit-full.html)---sceptical\n  of autonomous agents, built Solveit as the antidote\n- [Paige Bailey](https://softwareengineeringdaily.com/2025/01/09/ai-developer-tools-at-google-with-paige-bailey/)\n  on AI developer tools at Google and why Gemini 3 is built for \"acting and\n  coding\", not just chatting\n- [DHH](https://world.hey.com/dhh/promoting-ai-agents-3ee04945) flipped from\n  sceptic to enthusiast and made it look dramatic\n- [Donald Knuth](https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf)\n  opened with \"Shock! Shock!\" after Claude solved an open graph theory problem\n  he'd been working on for weeks\n- [Terry Tao](https://terrytao.wordpress.com/2025/11/05/mathematical-exploration-and-discovery-at-scale/)\n  on AI-assisted mathematical exploration at scale\n\nSo the issue isn't so much that there's _no_ theory for Agentic Coding, but\nthere are lots of nascent (and unverified) theories and it's hard to know which\nones are legit.\n\nBut I think this is an opportunity for my class; I'll have 100-200 (maybe more!\nestimating student numbers is hard) switched-on final-year and postgraduate\nstudents to try out these theories and see what works. One of the\n[weekly provocations](/blog/2026/02/20/comp4020-the-core-mechanic/) is\nexplicitly about that---finding one theory, using it to build a prototype, and\nreporting the results back to the class.\n\nWhat will we find? Who knows? The models will also be six months further on by\nthe end of the course, so the strengths/weaknesses and bottlenecks may shift\nfurther from where they are now. But I'll share the results of this experiment\non this blog---stay tuned.\n\n[^loudly]:\n    I mean, here I am\n    [writing a blog series about this course](/blog/tag/comp4020/), so I can't\n    exactly claim to be one of the quiet ones.\n",
  "createdAt": "2026-05-13T23:14:38.119Z",
  "description": "There are lots of nascent theories for agentic coding---here's a survey, and a plan to test them with 200 students.",
  "path": "/blog/2026/03/30/comp4020-whats-the-theory-here",
  "publishedAt": "2026-03-30T00:00:00.000Z",
  "site": "at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.publication/self",
  "tags": [
    "comp4020"
  ],
  "textContent": "There are lots of nascent theories for agentic coding---here's a survey, and a plan to test them with 200 students.",
  "title": "The road to COMP4020: what's the theory here?"
}