1 / 13

A Book by Rasmus Bornhoft Schlunsen

The Agentic
Crew

Engineering in the Age of AI Agents

March 2026

The Ground Is Shifting

"You write code, you ship it. That loop is breaking."
72% of developers used AI coding tools in 2025 Agents now write 30%+ of code at leading firms Old: write → run → ship. New: instruct → review → steer.

The Craft Is Molting

"The craft isn't dying — it's molting."

19 chapters covering:

Guardrails & Sandboxes Prompt Engineering Multi-Agent Orchestration Agents in CI/CD Testing as Feedback When NOT to Use Agents
A cocoon cracking open with light emerging

What Is an Agent?

The spectrum of AI assistance in software engineering:

Can read files, run commands, call APIs, and iterate on results. This is where most practical work happens today.
Rain Man — captain and savant illustration

The Rain Man Analogy

"You're Tom Cruise. The agent is Dustin Hoffman."

You — The Captain

  • Direction & vision
  • Judgment & taste
  • Knows when to walk away

Agent — The Savant

  • Speed & throughput
  • Tireless execution
  • Infinite pattern recognition

Three Core Capabilities

Planning

Click to reveal

Breaking down complex tasks into steps. Understanding dependencies, ordering work, and deciding what to do first.

Tool Use

Click to reveal

Reading files, running commands, calling APIs, writing code. The agent's hands in the real world.

Iteration

Click to reveal

Running code, seeing errors, fixing them, and trying again. The observe-adjust loop that makes agents powerful.

When Agents Fail

Scope Creep

Agent expands beyond the task, refactoring code you didn't ask it to touch.

Fix: Small, focused prompts. One task at a time.

Hallucinated APIs

Agent invents function signatures or library methods that don't exist.

Fix: Always run tests. Pin dependency docs in context.

Overconfidence

Agent presents wrong answers with certainty. No hedging, no doubt.

Fix: Verify everything. Trust but verify.

Context Loss

Agent forgets earlier instructions as the conversation grows long.

Fix: Fresh sessions. Project rules files. Keep context lean.

Storm clouds illustration

The Engineering Toolkit

Guardrails
Git
Sandboxes
Testing
Conventions
Click any node to learn more

Multi-Agent Orchestration in Practice

Wee — from theory to working code

Browser Nuxt 4 + Vue WSS Go Server 15MB binary REST + WebSocket SQLite embedded SDK Claude API + DeepSeek, GLM...
Multi-session = multi-agent orchestration (Ch 12)
Granular permissions = guardrails in code (Ch 4)
Pipeline integration = agents in CI/CD (Ch 13)
"The best way to learn agentic engineering is to build agentic tooling."
schlunsen/wee

Agents as Pentesters

What if your pipeline could hack itself before attackers do?

(Case study: Donna — open-source, Claude Agent SDK + Temporal)

Recon nmap, source analysis Analysis Injection | XSS Auth | AuthZ | SSRF 5 agents in parallel Exploitation Confirm vulnerabilities by executing attacks 5 agents in parallel Report CVSS, CWE, evidence Exploitation feedback loop Orchestrated by Temporal
"Every principle from the book — guardrails, sandboxes, orchestration, testing — converges here. If you can safely let agents attack your software, you've mastered agentic engineering."
schlunsen/donna

When Agents Attack (Responsibly)

"The same guardrails that protect you FROM rogue agents now need to protect the world FROM your intentionally hostile agents."

Isolate & Authorize

Sandbox your pentesting agents. Get written permission. Scanning production isn't testing — it's an incident.

Verify Everything

Agents hallucinate vulnerabilities. Every finding needs human eyes before it becomes an action item.

A ship with its crew

The Crew Metaphor

"You're the captain. The crew is disposable. The ship is not."
⚓ Hull = Your Codebase
⚖ Rigging = Conventions
📖 Logbook = Tests
♕ Captain = You
♟ Crew = Agents

Go Build Something

Get the Book -- Free