Zander Indo

An agentic-engineering system · built in the open

Atlas

Not one AI doing everything a little. A development team where every role is filled by a specialist — and one orchestrator that knows which specialist to call.

Hooks enforce. Agents judge. You decide.

1 orchestrator · 8 agent categories · 4 automation layers · 10 lessons

01 — The premise

One AI is a
generalist.

A single general-purpose assistant is good at a lot of things and great at nothing in particular. The moment you give it a clear role, a defined scope, and hard rules, it becomes dramatically more reliable. So I stopped treating it as one assistant and started building it as a team.

Every agent has a job it’s built for — and, just as importantly, jobs it’s not allowed to touch. Atlas sits on top, reads what you’re actually asking for, and routes it to the specialist that owns it. Named, if you’re wondering, after the Titan who holds up the world.

“Think of it less like one AI and more like a development team where every role is filled by a specialist.”

02 — The four automation layers

How it actually fires.

The most common wrong assumption is that agents lurk in the background and auto-fire on git events. They don’t. Agents are not daemons — they run in sessions you open on purpose. Automation is real, but it lives in four distinct layers, each with a different trigger and a different level of guarantee.

LAYER 01

Instructions

Always-on context that shapes every response. No trigger, no invocation — it just loads.

Passive
LAYER 02

Skills

Single-purpose capabilities invoked by name the moment a task matches their trigger.

Semi-automatic
LAYER 03

Agents

Specialists you open in a session, on purpose — orchestrated, sequenced, judged.

Orchestrated
LAYER 04

Hooks + CI/CD

Deterministic gatekeepers at the git boundary. They don’t judge — they enforce.

Mechanical

The AI is one layer — not the only layer. When CI catches something the agent missed, that’s also the system working.

03 — The team

Eight kinds of specialist.

Every agent falls into one of eight categories. It’s not decoration — the category decides where an agent sits in the pipeline and what it’s allowed to do. These are real, and they run.

01 · ORCHESTRATOR

Atlas

Reads intent, routes it to the right specialist. The one that holds the whole thing up.

02 · PLANNING

task-planner

Decomposes a ticket into a sequenced plan of agent invocations before any code is written.

03 · LEARNERS · RAG

context-reader

Read before you write — surfaces the naming and patterns already in the codebase first.

04 · DOERS

The builders

The agents that actually change code — refactors, scaffolds, migrations. Never unsupervised.

05 · FEEDBACK

code-reviewer

Naming, async correctness, edge cases, the smell you’d flag in a real senior review.

06 · GUARDIANS

security-auditor

The agents that can say no — OWASP, leaked secrets, accessibility gaps.

env-config-revieweraccessibility-auditor

07 · TOOL OPERATORS

GitHub MCP

The reach outside the repo — issues, PRs, and add-on servers via a single permission gate.

08 · PRESENTERS

implementation-summary

The AI work receipt — what changed, what still needs a human’s eyes.

pr-description-writer

Validators before doers. A doer without a validator is just an unreviewed factory.

04 — The pipeline

One command,
the whole gate.

A workflow agent is a fixed sequence baked into a single invocation. feature-pr-gate is the one I reach for before every pull request: you call the workflow, the workflow calls the specialists — in the order that matters.

STEP 01
code-reviewer
Correctness, naming, async, edge cases.
STEP 02 · PARALLEL
security-auditor + env-config-reviewer
OWASP surface and leaked-secret sweep, together.
STEP 03 · CONDITIONAL
dependency-auditor
Only if package files changed. Nothing runs for no reason.
STEP 04
pr-description-writer + implementation-summary
The PR body and the work receipt, written from the diff.

Different tools. Different jobs. Same pipeline. Because the order is the product.

05 — The pantheon

Named for Titans.
Built to grow.

The eight-category pattern isn’t specific to code. Every domain gets its own orchestrator, named for the Titan whose job it echoes — and its own color in the system. Atlas is the first seat, not the only one.

01 · DEVELOPMENT

Atlas

The Titan who holds up the world. Reads intent, routes it to the right specialist.

Live
02 · BUSINESS

Themis

Titaness of order & governance. User stories, INVEST audits, scope-creep, compliance.

Live
03 · DESIGN

Theia

Titaness of sight & radiant light. The design-system orchestrator — next to join the pantheon.

Next
04 · —

Open seat

Every new domain is one more Titan, one more color. The spectrum has room left in it.

The bridge between them

Titans don’t shout across domains — they hand off. Between them sits solution-bridge-writer: not a technical architect, a translator. It turns an approved product spec into a plain-language feasibility brief the dev side can act on. One handoff, fully traceable.

THE HANDOFF

Themis approves the spec
solution-bridge-writer translates it
Atlas plans & routes the build
feature-pr-gate guards the merge

Different Titans. Different domains. Same pipeline.

06 — What’s real, and what isn’t

No magic. On purpose.

Here’s the part most people selling this won’t say out loud: a prompt does not create a mind. There is no TechnicalWriter class behind the scenes, no behavior model. A well-written agent statistically steers output — it doesn’t guarantee it. So I draw a hard line between the two kinds of reliability.

MECHANICAL · 100% RELIABLE

Hooks & CI

A commit-message regex, a pre-commit lint gate, a CI run. These don’t have opinions. They either pass or they block — every single time. This is where you get real control: not vibes, actual constraints.

STATISTICAL · STEERED

Agents & prompts

Judgment, review, phrasing. Enormously useful, genuinely fallible. So they’re never the last line — a validator or a mechanical gate always sits downstream. The AI is one layer, remember.

That’s the whole reason I built it in the open and documented every decision: confidence should come from knowing where the guarantees actually live — not from believing the tool is smarter than it is.

07 — Learn the whole thing

Ten lessons.
Free. Start to team.

I didn’t just want faster output — I wanted to understand why it worked, well enough to teach it. So the entire system is a free curriculum: from a fresh install and your first orchestrator to a full team of specialists, workflow agents, and a business-side twin.

LESSONS 01–03

The mental model

Install, the context hierarchy, and the four automation layers — why agents don’t auto-fire.

LESSONS 04–07

The team & the gate

The eight categories, content types, building real agents, and automating GitFlow end to end.

LESSONS 08–10

Scale & the bridge

Skills & MCP, AI-first project setup, and the second orchestrator — Themis — on the business side.

Hooks enforce. Agents judge. You decide.