title: Agentic Engineering Practices — mvanhorn 完整工作流 title: Agentic Engineering Practices section: harness-engineering page_kind: concept sources: 1 status: candidate knowledge_status: ai_draft source_type: article judgment_owner: ai updated: 2026-06-03

Agentic Engineering Practices

What it is

A field-tested compilation of workflows, tools, and configuration patterns from three months of daily agentic engineering by @mvanhorn. The article documents 22 practical hacks that transformed the author from a non-shipping developer to a top open-source contributor across Python, Go, and major AI projects. The practices span planning, voice interaction, multi-session management, model routing, and real-world agent integration.

Why it matters

Most agentic engineering advice is either theoretical or tool-specific. This source provides a rare end-to-end operational blueprint: exact terminal configurations, plugin names, settings.json values, and day-in-the-life workflow descriptions. It demonstrates that the bottleneck is not model capability but harness design and operational discipline.

Key points

Plan-first workflow

• /ce-plan before any non-trivial task: The Compound Engineering plugin generates a structured plan.md with research, file targets, acceptance criteria, and pattern checks. For fuzzy ideas, use /ce-brainstorm first.
• Plans are for agents, not humans: The author skims the plan title and runs /ce-work. Inline questions ("TLDR?", "eli5 this plan") replace reading 300-line markdown files.
• Plan for the plan: For deep non-engineering work (strategy docs, competitive analysis), ask the agent to first plan how it will produce the deliverable, then execute that plan. This prevents corner-cutting.

Voice and input optimization

• Voice-to-LLM is different from voice-to-human: The transcription does not need to be perfect because the model fills gaps. Mumbling, trailing off, and restarting sentences are acceptable.
• Mac setup: Monologue or Wispr Flow pipes speech into the focused app. A gooseneck mic improves office usability.
• Phone setup: Apple built-in dictation is sufficient because the agent understands context even with mangled words.
• Raw transcripts into plans: Granola meeting transcripts are pasted directly into /ce-plan without summarization. The agent extracts relevant content against the codebase and prior strategy documents.

Multi-session operations

• 4–6 cmux tabs, each a separate Claude or Codex session: one writing a plan, one building from a different plan, one running a skill, one fixing a bug.
• New terminal tabs open directly into Claude Code: A launcher script (claude --dangerously-skip-permissions) removes the cd and claude typing friction. When a new session costs one keystroke, you start more of them.
• Remote control always on: Every session is reachable from the Claude mobile app. Start at your desk, pick up the same live run on your phone.

Permissions and automation

• skipDangerousModePermissionPrompt: true with an allowlist in ~/.claude/settings.json: With six sessions, babysitting permissions is not viable. The author recommends bypass mode with a sound hook on session stop.
• AgentMail email address: A daemon watches an AgentMail inbox over WebSocket. On every allowlisted email, it opens a fresh Claude session and acts on the subject and body. Bug reports emailed from a phone become working sessions before returning to a screen.

Model routing and dual-engine setup

• Claude plans, Codex builds: The author delegates construction to Codex without leaving the Claude session. Three handoff paths: Codex IDE extension, /ce-work --codex, and Printing Press Codex mode.
• Settings: Codex runs at reasoning xhigh with fast mode on; Claude Code runs at xhigh with fast mode off (to avoid per-token billing on the Max plan).
• Two $200 plans side by side: Parallel builds go to Codex; planning and taste stay on Claude. Some friends reverse the roles.

Research before planning

• /last30days before /ce-plan: The last30days skill searches Reddit, X, YouTube, HN, TikTok, Instagram, Polymarket, GitHub, and the web in parallel. Research, plan, build is the real loop.
• Example: Choosing between Vercel agent-browser and Playwright. last30days surfaced that agent-browser uses far less context per call; Playwright dumps thousands of tokens for tool definitions. The plan came out grounded in community knowledge, not training data.

Skills and compounding

• Anything done more than twice becomes a skill: Reusable commands that persist across sessions. The trick is pointing the agent at a working skill and asking it to copy the shape.
• Skill editor and no-ai-slop skills: A /skill-editor skill keeps all skills concise. A /no-ai-slop skill strips em dashes, "X, not Y" phrasing, and duplicate instructions.
• Open-source contributions: The same loop ships personal projects and external ones. The author has had hundreds of PRs merged into Python, Go, OpenCV, Vercel Agent Browser, and OpenClaw.

Real-world agent integration

• Printing Press: A fleet of CLIs that wrap real-world services (Tesla, Instacart, ESPN, Alaska Airlines) so an agent can execute errands. Agent Cookie provides browser session auth without re-authentication.
• HyperFrames: Video built as HTML — an agent writes the script, scene by scene, and renders to MP4. The cost of a video dropped to a conversation.
• Notes as knowledge base: Bear CLI (decade of notes), Obsidian, gbrain, and supermemory pointed at agents so personal knowledge compounds.

Human role redefinition

• Be the signal, not the hands: When running six agents, the human job is taste, direction, and the react-and-redirect loop. Agents supply volume; you supply judgment.
• Agent psychosis warning: Building with agents is addictive. The author warns about friends who "are so lit up by being able to build anything that they do not do anything else." The trap is not empty launches; it is vanishing into the build and losing people around you.

Evidence across sources

Open questions

• How many of these hacks are Claude Code-specific versus transferable to Codex, OpenClaw, or other agent platforms?
• Does the multi-tab approach scale beyond 6 sessions, or does cognitive overhead plateau?
• What is the maintenance burden of Printing Press CLIs when third-party services change their APIs or auth flows?
• How does the "be the signal" role change as agents improve at taste and direction?

Prompts for witness

• Which of the 22 hacks would have the biggest impact on your current workflow if implemented today?
• You run multiple agent sessions already. At what point do you feel the cognitive overhead of context-switching exceeds the benefit?
• The article warns about "agent psychosis" — building so much you disconnect from people. Where is your own boundary between productive building and compulsive building?

Related

• harness-engineering/compound-engineering — The planning framework that underpins many of these hacks
• harness-engineering/skill-engineering-as-algorithm — Architectural theory behind skill design
• harness-engineering/barebones-agentic-engineering — Counterpoint: minimal tooling vs. mvanhorn's comprehensive stack
• claude-code/overview — Claude Code platform details
• harness-engineering/multi-agent-coordination-patterns — When one session becomes many
• harness-engineering/headless-mac-mini-setup — Remote agent infrastructure