Agentic Engineering: The Guide After Vibe Coding (With Real Numbers)
Karpathy says vibe coding is over. Here's what agentic engineering actually means in practice, with the cost numbers and architecture choices that prove it.
Research foundations.
The state of the art behind our engineering.
Academic papers, theoretical frameworks and conceptual foundations that ground how Swoft builds AI agents and software architecture: Sheridan-Verplank autonomy levels, neurosymbolic engineering (Mastropaolo & Poshyvanyk FSE 2025), agent definitions (Russell-Wooldridge-Ferber), Conway's law, MCP, Deep Tech (BCG-BPI-Hello Tomorrow).
Agentic Engineering: The Guide After Vibe Coding (With Real Numbers) What is an AI agent? Returning to 70 years of academic research What is an AI agent? Returning to 70 years of academic research
ChatGPT isn't an agent in the academic sense. Russell, Norvig, Wooldridge and Ferber have proposed for 30 years a demanding definition that most 2026 products don't meet. State of the question.
Multi-agent systems: from Ferber to executable Conway's Law Multi-agent systems: from Ferber to executable Conway's Law
Jacques Ferber wrote Multi-Agent Systems in 1995. Thirty years later, his organizational dimension — often forgotten by contemporary frameworks — becomes production-ready thanks to event-driven architecture.
What is Deep Tech? A rigorous definition and the fault lines What is Deep Tech? A rigorous definition and the fault lines
The term Deep Tech has covered everything and its opposite since 2014 — quantum, biotech, fusion, AI, blockchain. Let's go back to the canonical definitions of BCG, Hello Tomorrow and Bpifrance to distinguish what is from what isn't.
French Deep Tech in 2026: where the ecosystem stands French Deep Tech in 2026: where the ecosystem stands
Mistral, Pasqal, Aledia, Verkor, Diabeloop: France has more Deep Tech than most European countries but struggles to turn these gems into global champions. State of the ecosystem in 2026.
The 5 autonomy levels of an AI agent — and why 90% of projects stop at level 2 The 5 autonomy levels of an AI agent — and why 90% of projects stop at level 2
The Sheridan framework adapted to 2026 business: five tiers from simple assistant to fully autonomous system. Diagnosis, architectural requirements and a realistic timeline to move up a level.
Neurosymbolic architecture: the way out of the parrot-versus-solver compromise Neurosymbolic architecture: the way out of the parrot-versus-solver compromise
How to combine LLMs and formal reasoning to build reliable AI agents on specialized domains. A reading of Kautz's taxonomy and the AlphaGeometry, Plato-3, FAOS cases.
Why real AI agents in production are so rare in 2026 Why real AI agents in production are so rare in 2026
Technical diagnosis of the four structural obstacles that block the transition from demo to real production. And what it takes to clear them.
MCP, the HTTP of AI agents, explained for CTOs in 2026 MCP, the HTTP of AI agents, explained for CTOs in 2026
The Model Context Protocol published by Anthropic in 2024 has become the de facto standard to connect LLMs to tools. Why it matters, how to adopt it, what to watch out for.
Multi-agent coordination: 14 failure modes and how to avoid them Multi-agent coordination: 14 failure modes and how to avoid them
The Cemri 2025 paper identified 14 recurring failure modes in multi-agent systems. Diagnosis and three architecture families to guard against them.
Other reading registers
Explore the rest of the editorial library
A paper to discuss?
We read everything serious that gets published.
Swoft's R&D team monitors academic literature on AI agents, neurosymbolic engineering and software architecture. If a paper raises a question for you, we're happy to dig into it together.