RaduWeb.

System Philosophy

Agentic Engineering &
The Right Way to Build

We don't talk about fake projects. Here is our architectural philosophy on how to orchestrate frontier AI agents, avoid auto-pilot loops, and build high-reliability software.

Red Book representing documentation archives and engineering philosophy
Structured architectural workflow design specs

The Fallacy of Full Autopilot

Dumping an entire project onto an agent and running it on auto-pilot inevitably leads to compilation loops, runaway token bills, and code rot. High-fidelity software requires strict verification gates.

Verification GatesHuman ApprovalStructured PlansContext Control

Autopilot Pitfalls

Agents run in unmonitored cycles trying to brute-force solve bugs. Without boundaries, they edit unrelated components, introduce silent regressions, and exhaust rate limits on minor issues.

The Correct Way

Every action is partitioned: a research phase creates an explicit Implementation Plan. The agent must pause and obtain approval before executing commands, compiling continuously at every step.

Red Airplane representing parallel agentic flight paths and isolated context routes

Decomposed Context vs. Monoliths

Dumping a million tokens of files into a prompt generates noise and retrieval errors. Precision engineering requires breaking the workspace down into narrow, isolated agent contexts.

Subagent DispatchTask ScopingSandboxed RuntimesParallel Orchestration

Monolithic Degradation

Loading the entire codebase into a single workspace causes the LLM to get lost in the details. The agent loses track of the core system rules, hallucinating imports and function signatures.

Parallel Subagents

We dispatch narrow, specialized subagents for distinct goals (e.g. styled frontend tweaks, database checks, API testing) operating in parallel, and merging their outputs into a central verified branch.

Stoplight representing diagnostic gates and compiler safety checks

Deterministic Safety over Raw Shells

Executing unvalidated terminal commands or running global scripts breaks system security and ruins development sandboxes. Tools must be sandboxed and type-safe.

Type ValidationMCP Tool IsolationAST ParsingDependency Protection

Shell Exposure

Allowing agents to run raw bash commands or global `pip install` risks data loss, environment corruption, and security breaches. There is zero structural guarantees on the outcome.

Safety-Guided MCP

Agents utilize typed schemas (Model Context Protocol), AST-based partial editors, local virtual environments, and continuous static lint verification to ensure code is compile-safe at every point.

Flight log // shipped

Selected work

Sites I have built and shipped for clients in Romania, the UK, Spain and South Africa, from a university's online degree platform to e-commerce, venues and viral campaigns.

Higher education

University of Hertfordshire Online

The university's platform for 100% online master's degrees: programme pages, a course-matching quiz and the route from enquiry to application.

  • Webflow
  • Lead funnel
  • Programme catalogue

online.herts.ac.uk ↗

Beauty e-commerce · Romania

Tuju.ro

A Romanian beauty store: product catalogue, promotions and loyalty, checkout, plus analytics and consent tracking wired correctly.

  • Shopify
  • E-commerce
  • Tracking & consent

tuju.ro ↗

Event venue · Bucharest

Casa Paulownia

Website for a Bucharest event venue: spaces, weddings and private events, and an enquiry path for bookings. Fast static pages with structured data for search and AI assistants.

  • Static / SSG
  • Local SEO
  • AI-readable

casapaulownia.ro ↗

Campaign microsite

HiSmile Romania

A campaign microsite where visitors capture their smile in a branded frame and enter a daily prize draw. Built to handle a viral traffic spike.

  • Camera capture
  • Campaign
  • High traffic

hismileromania.ro ↗

More clients // manifest

10 entries · and others