Agentic Systems Engineer

Erik Hill

I build agentic systems that check their own work — and keep a human on every step that can't be undone.

Self-taught in four months. I designed and built a full autonomous development harness: a multi-agent loop that plans, executes, critiques, and validates its own changes to a shipping React / Capacitor product — with the guardrails to trust it.

// Pipeline

A loop that reviews itself before it ships

Work moves through five stages. Each hands off to the next automatically — no manual relaunch — and judgment runs on a different, colder model than execution, so the critic isn't grading its own homework.

Per-role model routingReasoning-heavy judgment and routine execution run on different models, matched to the job.
Hook-based handoffStages arm the next leg on their own — the loop advances without a human babysitting it.
Cold, independent criticReview runs in fresh context so it catches what the builder rationalized away.
// The Oracle

Correctness it can prove — not vibe-check

The same game logic runs in two independent implementations — a React build and a native engine. Logic-invariant tests hold both to the same rules, so a regression in one is caught by the other before it ships.

React implementation ≟ must agree ≟ Native engine differential oracle

On top of the differential check: invariant enforcement on the core game rules, plus on-device validation driven by adb automation — so "it passed" means it passed on real hardware, not just in a mock.

// Autonomy Ladder

Automate the launches. Never the approval.

Trust is earned in named steps. Higher rungs remove clicks, never gates — a human still approves anything irreversible or outward-facing (deploys, commits, anything a user would see).

L4Lights-out — a doctrine amendment, not a default. Operator's call, never self-granted.gated
L3Auto-advance launches — the safe clicks fire on their own; write-gates still stop for a human.gated
L2Hands-off relay — the loop chains itself end to end; every irreversible act still waits for approval.◀ current
L1Assisted — humans drive; the harness proposes and prepares each step.manual
L0Manual — every action is a human keystroke.manual

Rule: reliability earns launch-automation — it never earns gate-removal.

// Stack & Practice

What I actually work in

Agentic

  • Multi-agent orchestration
  • Prompt & context engineering
  • Evals & oracle design
  • Human-in-the-loop safety

Build

  • Python
  • JavaScript / React
  • Capacitor · Android (adb)
  • Git

Systems

  • Capability catalogs
  • Validation pipelines
  • Device-test automation
  • Governance / doctrine
// The Path

Not the usual route

I came to this from professional kitchens, not a CS program — six years of running high-volume lines, then four months of building agentic systems from first principles. Judge me on the work, and I'll take that trade every time.