> Source URL: /openai.casestudy # OpenAI Build Week 2026 — PathMX Learning Labs **Event:** [OpenAI Build Week 2026](https://github.com/pathmx/pathmx-build-week-2026) · [Devpost submission: Learn with Codex](https://build-week.pathmx.net/) **Team:** Mark Johnson, Tram Le, Andrew Miller **Role:** Learner testing + research **Tools:** Codex / GPT-5.6, PathMX, Markdown-first curriculum Sources, eval harness --- # Summary [PathMX Learning Labs](https://build-week.pathmx.net/) asks a simple question: what if a useful conversation with an agent became a learning path you could keep, change, and continue tomorrow? Humans working with AI agents often produce remarkable explanations—then that work vanishes into chat history. Real learning needs more than an answer: a durable goal, an intentional sequence, practice, interactive representations, evidence of progress, and a useful next step. Our submission, **Learn with Codex**, treats Codex as the conversation and orchestration surface and PathMX as the durable, playable learning surface. A learner starts with one prompt and a personal goal. Codex creates a local learning repository, asks about the starting point and constraints, and proposes a milestone map before building the first lesson. The chat becomes orchestration; the generated Path becomes the lasting product. I worked on the team as a beginner-facing tester and research contributor: walking the early learner loop as someone who did not already know PathMX, contributing the Campus Constellation exploration, and surfacing structure and waiting-time problems that shaped the buffered `/learn` workflow. --- # Inspiration - [PathMX methodology](https://pathmx.dev/index.path) — curriculum as readable, executable Markdown Sources - [Matt Pocock’s `/teach` skill](https://www.aihero.dev/learn-anything-with-my-teach-skill) — portable personal learning with an agent - A portable executive curriculum designed for personal learning goals - [Google’s Open Knowledge Format](https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing) — durable, shareable knowledge structures - Dogfooding the methodology: we used Codex to build the learning system we were investigating --- # Problem Coding agents can teach in the moment, but most of that help disappears when the chat ends. For nontechnical learners especially, three gaps show up quickly: 1. **No durable artifact** — useful explanations stay trapped in a transcript. 2. **No visible structure** — the agent may teach before the learner has confirmed a path. 3. **Long silent waits** — a strong final module still feels broken if the learner sits through multi-minute turns with no progress. Waiting time is not only a latency metric; it is attention friction. While an agent builds course materials, a distracted human can wander—and the “live” learning experience fails even when the final files are correct. The research question we treated as first-class: starting from only a hosted instruction file, can a coding agent create and maintain a useful personal learning repository for a nontechnical learner? --- # My Role - Tested the early learner loop as a beginner, not as someone who already knew the toolchain. - Contributed the [Campus Constellation](./campus-constellation-networking.path.md) exploration as a concrete learning experiment ([playable lab](https://github.com/pathmx/pathmx-build-week-2026/blob/main/paths/labs/campus-constellation/index.demo.md)). - Flagged structure problems: when the agent taught before a map existed, or left the plan only in chat. - Flagged waiting-time problems: long silent turns that made the experience feel unusable even when the final artifact was good. - Helped turn those observations into product pressure for a buffered `/learn` workflow—early persisted maps and staged modules so learners have something to open while generation continues. - Reviewed experiences as a learner and helped decide which results were worth preserving in the durable repo. --- # What We Built A document tells you something. An app lets you do something. A PathMX Path can guide you through understanding, doing, deciding, and creating—and it remains yours when the session ends. The public submission has cooperating pieces: | Artifact | Responsibility | | --- | --- | | Hosted bootstrap | One prompt and one stable URL for a nontechnical learner | | PathMX Skills | `/pathmx` (authoring), `/learn` (personal learning), `/teach` (shared curricula) | | Learning Starter | Milestone maps, module scaffolds, Player tutorial, durable profile, validity checks | | PathMX Player | Turns those Sources into focused, interactive experiences | | Eval harness | Real Codex CLI through a five-phase learner journey, scored on artifacts, quality, and latency | We built *Learn with Codex* with Codex itself. The repository was the project manager: typed tasks with human owners, a playable change log agents must update before every push (enforced by a check script), research briefs, and evaluation evidence—all ordinary Markdown that humans and agents read from the same context. PathMX was created by Mark Johnson before Build Week. During Build Week, the team used Codex and GPT-5.6 to extend that foundation with teaching skills, starter materials, interactive labs, reusable components, evaluation evidence, and a durable human-agent research workflow. Humans selected the questions, reviewed the experiences as learners, and decided which results were worth preserving. Public companions: - [PathMX Build Week 2026](https://github.com/pathmx/pathmx-build-week-2026) — labs, tasks, and the living record of collaboration - [PathMX Learning Starter](https://github.com/pathmx/pathmx-learning-starter) — learner-facing starter for personal Paths - [PathMX Skills](https://github.com/pathmx/pathmx-skills) — canonical agent skills (`bootstrap.md` is the one-prompt entry point) - [Learning Labs site](https://build-week.pathmx.net/) — playable submission surface generated from the same Sources --- # Approach ## 1. One-prompt entry Learners open an empty folder in Codex Desktop and send: ```text Follow the instructions at https://raw.githubusercontent.com/pathmx/pathmx-skills/main/bootstrap.md. Create a new learning space in ./learning-space and help me learn [your topic or goal]. ``` The agent installs tools, creates the learning space from the Starter, asks short learner questions, and opens the first playable map. The primary journey requires no API keys, plugin development, or manual PathMX setup—any human (or their mom) with Codex Desktop should be able to start. ## 2. Keep humans and agents in the same repository Work, decisions, and agent instructions lived in the Build Week repo so context did not have to be reconstructed from chat: - Tasks — Codex-managed work queue and evidence trail - Labs — experiments that may fail (16+ playable demos) - Research — external references and the learning-agent eval brief - Work log — dated notes, handoffs, and a playable change log required before shared pushes ## 3. Evaluate the full learner journey We did not score a single clever reply. The harness ran real Codex CLI sessions (0.142.5–0.145.0) through onboard → proposed map → confirmed map → complete module → learner return, and measured five acceptance dimensions: 1. **Structure** — persist a whole-path map before substantial teaching 2. **Learning quality** — examples, practice, hints, review, and checks 3. **Repository quality** — valid PathMX and durable progress evidence 4. **Continuity** — honest adaptation when the learner returns 5. **Responsiveness** — useful progress without long silent waits between lesson steps Latency was scored separately from quality on purpose: artifact quality is not learning quality. Pacing, focus, practice, and feedback matter as much as correctness. --- # Challenges Our biggest constraint was too many promising ideas for the time available. Early attempts to generate a complete “learn anything” product exposed several problems: - Interactive components needed careful lifecycle, focus, and navigation behavior. - Feature selection for better learner UI/UX was hard under a short timeline. - **Waiting time caused real friction** — the minutes an agent spent building course materials were minutes a distracted learner could wander. That observation from beginner testing became a first-class product requirement, not a nice-to-have. --- # What the Evidence Showed Early default-strength runs could produce valid repositories and still fail as learning experiences. One ambiguous-goal run passed artifact checks while its final turn took **12m43s**. Early runs stretched as long as **18m55s**. That kind of wait is exactly what beginner testing made hard to ignore. Failure-driven instruction changes followed a clear pattern: | Observed failure | Resulting change | | --- | --- | | Module authored before map confirmation | Map-first phase contract | | Map left only in chat | Persist and link the proposed Path | | Valid files without enough learning support | Required example, hint, feedback, review, checkpoint | | Long silent work | Early map, staged updates, buffered module | **Final candidate flows** scored **100% deterministic + 100% judge** at **6m56s** and **10m07s**, with a first useful update in **4–13 seconds** on every measured turn. Buffered modules and early persisted maps became the primary learner-speed strategy—the same direction my waiting-time and structure feedback pushed toward. Other accomplishments from the submission: - A one-prompt learner flow a nontechnical person can run - **16+ playable lab demos** — from a Kepler orbit instrument and interactive chess to a practice-interview voice simulator, a research bench, Campus Constellation, and personal-knowledge experiments - A public Build Week Space at [build-week.pathmx.net](https://build-week.pathmx.net/) where the submission deck itself is a playable, cloneable Path - A human-agent workflow whose tasks, decisions, evals, and changes remain inspectable as plain Markdown Details live in the team’s [learning-agent eval brief](https://github.com/pathmx/pathmx-build-week-2026/blob/main/paths/research/learning-agent-evals.brief.md). --- # What We Learned - **Durability changes the agent relationship.** A shared file workspace lets learning and decisions compound across sessions. - **Artifact quality is not learning quality.** Responsiveness had to become its own scored dimension. - **Repository contracts beat prompting.** Encoding rules in skills, templates, and checks means the next agent recovers correct behavior from the repo, not hidden context. - **Learner actions should become evidence.** Durable responses give the learner and the agent a shared basis for what comes next. - **The process can embody the thesis.** Our own work with Codex got more coherent when research, tasks, and reviews lived in the same durable graph. - **Beginner testing is evidence.** If a nontechnical learner cannot see progress, the system is not finished—even when the final Markdown is correct. Codex and GPT-5.6 accelerated research, authoring, verification, and synthesis. The team still chose direction, set learning goals, reviewed experiences as learners, and accepted or rejected results. --- # Links - [Public Learning Labs](https://build-week.pathmx.net/) - [Build Week repository](https://github.com/pathmx/pathmx-build-week-2026) - [Campus Constellation research](./campus-constellation-networking.path.md) - [Campus Constellation lab](https://github.com/pathmx/pathmx-build-week-2026/blob/main/paths/labs/campus-constellation/index.demo.md) - [PathMX Learning Starter](https://github.com/pathmx/pathmx-learning-starter) - [PathMX Skills](https://github.com/pathmx/pathmx-skills) - [PathMX methodology](https://pathmx.dev/index.path) --- # Next Steps - Validate collaboration properly: rerun larger modules with subagents disabled vs. required before claiming any speed effect - Run the manual Codex Desktop journey to measure permissions, Browser handoff, and real-session latency - Ship second-wave labs (negotiation rehearsal, debugging dojo) and graduate reusable learning patterns into the Starter - Explore stateful and multi-user Paths once the single-player loop is boringly reliable - Open-source the pieces of PathMX that can be separated from Core as its boundaries stabilize - More beginner-facing labs that stress structure and wait-time the way [Campus Constellation](./campus-constellation-networking.path.md) did --- ## Backlinks The following sources link to this document: - [Open case study](/index.path.llm.md) - [Open case study](/case-studies.index.llm.md) - [OpenAI Build Week case study](/campus-constellation-networking.path.llm.md)