Mid-2026 field guide · updated continuously

Professional Agentic Product Engineering Guide (Mid-2026 and updated continuously)

Main maintainer: Alexey Krivitsky (alexey@krivitsky.com)
Upstream repo: https://github.com/krivitsky/professional-agentic-product-engineering
⭐ Star it; submit issues or PRs — help yourself and the next person learn better.

Goal of this Guide

Getting good at operating a coding agent (using the example of a popular agentic coding harness, Claude Code by Anthropic) for creating new software and working on real codebases.

It spans the full range: from "fix bug xyz" all the way to autonomous engineering loops running in production.

Calibrated for the current frontier class — Opus 4.8+, GPT-5.5-class+, Gemini 3.x+.

The one idea

Professional agentic engineering is not prompt engineering. It's engineering the system around the model.

As the work gets harder, where you apply effort climbs a ladder — the prompt shrinks while the system around it grows:

flowchart LR
  P[Prompt] --> T[Task] --> C[Context] --> V[Verification] --> E[Environment] --> X[Execution]

The eight tiers below are the detailed rungs of that one climb (Prompt = T1, Task = T2, Context = T3, Verification = T4, Environment = T5–T7, Execution = T8). Every tip is an instance of one layer — learn the ladder and the 60 tips fall into place.

Who this is for

Engineers and technical founders — operate an agent in a real repo, not vibe-code a demo.
Product managers closing the tech gap — ship real changes, not just specs.
Senior leaders who want real hands-on experience, not slideware.
Non-IT professionals entering product development in the age of AI.

Your level — where to start

If you've used a coding agent a few times and want to get professional, start at the top and stop climbing wherever you are today.

Already more fluent? Jump straight to the tier that matches you — or tell your agent to skip ahead to the sections you need.

TL;DR — if you do only five things

Be specific and positive. Name files, constraints, and the pattern to follow; say what to do, not what to avoid. (Tips 1–4)
Give the agent an executable Definition of Done. Tests/lint/typecheck as commands — that's the loop's exit condition. (Tip 31)
Plan before you edit; slice the work thin. Investigate → approve a plan → one vertical slice at a time. (Tips 15, 17, 19)
Commit on every green step. Each commit is a checkpoint the loop can revert to. (Tip 40)
Engineer the environment, not the prompt. CLAUDE.md, Skills, hooks, MCP, CI carry the intelligence. (Tip 51)

The eight tiers at a glance

Tier	You learn to…
T1 Professional Prompting	Write prompts the agent can act on
T2 Planning & Slicing	Plan and slice before you build
T3 Context Management	Give the agent the right context and tools
T4 Loop Until Done	Make the agent prove it's done (the heart of it)
T5 Checkpointing & Hardening	Checkpoint in git; wire tests & CI into the harness
T6 Orchestration	Run many agents at once
T7 Fleet Ops	Operate your agents as a fleet
T8 Agent Execution Layer	Put agents into production (the execution layer)

Climb only as high as your work demands — then stop.

Climb the eight tiers

Each tier is its own chapter — open the one your work needs.

T1Professional Promptingthe agent keeps doing almost the right thing — vague asks get literal, wrong results. The single request is your cheapest lever: say what you want so it can't guess wrong.T2Planning & Slicingbig asks go sideways — the agent edits the wrong things or tries to do everything in one pass. Plan first, then cut the work into small slices, each one small enough to check and cheap to undo.T3Context Managementyou re-explain the same conventions every session, or the agent can't see your DB, browser, or docs. Engineer that context once — durable project memory plus the right tools — and it stops re-asking and guessing.T4Loop Until Doneyou can't trust the output without reading every line, and "done" means nothing concrete.T5Checkpointing & Hardeninga long run goes wrong and you've lost good work with nothing to roll back to. With a checkpoint at every green step, a bad run costs minutes to undo, not a rewrite.T6Orchestrationone agent is too slow or floods its own context, and the build is too big for one pass. The next gain is parallelism — many subagents on independent slices, specialist roles, and a second model catching what the first missed.T7Fleet Opsruns die because your laptop slept, parallel agents collide, or you want to drive them from your phone. Get this right and a run survives your laptop closing and keeps going while you're away.T8Agent Execution Layerthe team needs it — agents must pick up tickets and open PRs without anyone babysitting a terminal.