Agentic Coding with OpenAI Codex CLI Build intelligent agent workflows using Agentic Engineering, MCP, hooks, and delivery… (Daniel Vaughan)（Z-Library）

【One-Line Pitch】 A practical field manual for developers who want to move beyond autocomplete and run OpenAI's Codex CLI as a real agent inside their repo, tests, and CI—covering prompting, sandboxing, orchestration, and team rollout. Best for terminal-comfortable engineers and leads who need a safe adoption framework. 【Book Arc】 - **Opening (~0%–10%)**: Frames the core distinction between reactive text-completion tools and Codex CLI's agent loop (plan, execute, observe, adjust), and situates it among IDE, terminal, and cloud agents. Solves the "why this is different" question. - **Early (~10%–32%)**: Gets you running—setup, authentication, the quickstart, and the discipline of reading test results rather than keystrokes. Introduces prompting with goal/context/constraints/completion criteria, AGENTS.md rules, plan mode, and PLANS.md for long tasks. - **Early–Middle (~29%–42%)**: Establishes the trust model: approval modes and kernel-level sandboxing as two independent controls, then model selection and reasoning effort as the capability/cost levers. Moves into context and cost management, including sub-agent delegation as a context technique. - **Middle (~42%–55%)**: Extends the agent's reach—MCP servers (including building a minimal one with FastMCP), hooks that intercept the agent lifecycle, skills, and sub-agent orchestration with worktrees. - **Late (~55%–75%)**: Operationalizes agentic work: debugging and testing agentic workflows, failure taxonomies, AI code review with the /review command and PR integration, and CI/CD delivery automation. - **Ending (~75%–100%)**: Widens to practice and strategy—practitioner guides (migration, backend, frontend, infrastructure as code), benchmarks and competing tools, harness engineering, and how to structure an agentic engineering team. 【Key Takeaways】 - **The agent loop, not autocomplete, is the unit of value** (Opening): Codex plans, executes, observes, and adjusts, so it excels at tasks with a beginning, middle, and end—refactors, end-to-end features, failing-build diagnosis—rather than line-by-line suggestions. - **Fast, reliable tests are the highest-leverage asset in your repo** (Early): The agent verifies by running tests, so slow or flaky suites collapse its feedback loop and turn first guesses into final answers. Watch test results, not intermediate edits. - **Approval and sandboxing are separate controls** (Early–Middle): `--ask-for-approval` governs when a human is asked; `--sandbox` governs what the OS permits. Combining them deliberately (untrusted/read-only for audits, never/workspace-write for CI) is the core safety skill. - **Model choice sets the ceiling; reasoning effort sets how close you get** (Middle): Pin models and effort in config.toml and named profiles, and allocate reasoning budget where it pays—flagship models on trivial lookups waste quota, lightweight models on architecture quietly cost more later. - **Delegation is context management** (Middle): Sub-agents run in fresh windows and return compact summaries, so the parent keeps outcomes rather than transcripts. Delegate only past a threshold (several files, many commands, multiple fix cycles) or overhead rivals inline cost. - **MCP, hooks, and skills turn the CLI into a platform** (Middle): MCP servers add tools (and can be built quickly with FastMCP), while lifecycle hooks (PreToolUse/PostToolUse) enable auditing, context injection, and auto-testing—even intercepting MCP calls. - **Team memory is a real gap you must engineer around** (Middle): Native memory is single-user; the practical workaround is a git-backed shared knowledge repo wired through SessionStart/Stop hooks. - **Agentic delivery needs its own review and testing discipline** (Late): AI code review, PR integration, failure taxonomies, and benchmarks like c-CRAB and the over-mocking problem give you calibration tools rather than blind trust. 【Reading Tips】 - Deep-read the trust-model and context/cost chapters (approval modes, sandboxing, model effort, delegation)—these are the decisions that determine safety and spend; skim the landscape/competitor survey if you already know the tooling. - Treat the exercises as the point: writing a four-element prompt, auditing AGENTS.md, and running plan mode on a cross-cutting change will teach more than passive reading. - Before adopting at team scale, jump to the hooks, team-memory workaround, and CI/CD sections—they address the gaps solo usage hides. - Keep the failure taxonomy and benchmark chapters as a checklist when your agent sessions go wrong, rather than reading them linearly. 【Coverage Limits】 This guide is synthesized from stratified excerpts and the book's front matter; specific chapter-level detail in later practitioner guides and enterprise deployment is only partially represented, so treat those sections as directional rather than exhaustive.