Codex in April 2026 is a layered system: AGENTS.md for durable policy, skills for repeatable procedures, hooks for deterministic enforcement, subagents for context isolation, Codex Cloud for long-running work, MCP for external context, computer use in GPT-5.5 for UI work, plugin marketplaces for distribution, and Codex Security for vulnerability scanning. Most teams use two or three of these. The cost of skipping the rest shows up later as review debt and invisible changes. This course teaches the design decisions behind each layer and how to wire them together.
Course objectives
Target audienceSenior engineers, staff and principal engineers, tech leads, platform and DevEx teams, internal AI champions. The work participants leave with is the work their team will inherit.
Prerequisites
Module 1.
Codex as a layered system. AGENTS.md for policy. Skills for procedures. Hooks for enforcement. Subagents for context isolation. Codex Cloud for long-running work. MCP for external context. Computer use for UI tasks. Participants receive the Codex Feature Matrix as a day-one handout.
Module 2.
MCP and real context without new risk. Where external context changes outcomes (issue trackers, error monitoring, production logs). Where it just expands attack surface. Trust boundaries: which servers can read code, write code, hit production. Lab: select three MCP servers and justify each against a risk checklist.
Module 3.
Skills, plugins, and team assets. Turn repeated work into reusable skills. Bundle related skills, hooks, and subagent definitions as a plugin so a new hire inherits team conventions on day one. Lab: author a skill from an existing repeated workflow and publish it internally.Module 4. Codex Cloud, subagents, and computer use. Three primitives that solve different problems. Codex Cloud for long-horizon tasks (8-hour migrations, cross-repo refactors). Subagents for fire-and-forget context isolation. Computer use (GPT-5.4 and later) for UI work where the agent needs to see and click. Lab: same task implemented three ways, compared on cost, time, correctness.
Module 5.
Hooks, sandbox modes, approval policies. Hooks are the deterministic layer. Sandbox modes (read-only, workspace-write, danger-full-access) define what is technically possible. Approval policies (untrusted, on-failure, on-request, never) define when the agent has to ask. The dangerous combination is danger-full-access plus never on a developer laptop. Lab: write three hooks for the participant's own repo, document a permission policy.
Module 6.
Capstone. Each participant designs a 90-day rollout with three phases (pilot, expansion, standard practice), named owners, success criteria, and a kill switch. The plan must include at least one headless pattern (codex exec) running in CI. Plans defended to the room.
Duration: 1 day
Price: 10 900 NOK
Language: English
Format: Classroom, virtual classroom, or in-company
Hva er Codex Advanced?
Codex Advanced er et videregående kurs for erfarne brukere som ønsker å jobbe strategisk og strukturert med Codex i større utviklingsprosjekter og team. Kurset fokuserer på arkitektur, kontroll og skalerbar bruk av AI i utvikling.
Hvem passer kurset for?
Kurset passer for seniorutviklere, tech leads og andre tekniske nøkkelpersoner som har erfaring med Codex og ønsker å bruke det mer systematisk i team og organisasjon.
Trenger jeg forkunnskaper for å delta?
Ja, kurset forutsetter erfaring med Codex i praksis, samt god forståelse av utviklingsprosesser og teamstandarder.
Er kurset praktisk rettet?
Ja, kurset jobber med realistiske scenarier og fokuserer på hvordan Codex brukes i produksjonssetting, inkludert struktur, kontroll og kvalitetssikring.
Hva lærer jeg som er nyttig i praksis?
Du lærer blant annet hvordan du designer en lagdelt Codex-setup, bruker subagenter og agentteam, setter riktige tilgangsnivåer, bygger gjenbrukbare komponenter og planlegger utrulling i team.
