AIE Europe Day 2: ft Google Deepmind, Anthropic, Cursor, Factory, Linear, HF, Cerebras & more
Summary
The AI Engineer Europe conference featured several key discussions on the evolving landscape of AI and software development. Google DeepMind announced Gemma 4, a family of open models ranging from 2B to 32B parameters, capable of running on devices from Android phones to consumer GPUs, with multimodal understanding and an Apache 2 license. Anthropic highlighted the rapid growth of the Monocontext Protocol (MCP), reaching 110 million monthly downloads, and discussed its role in general agents for knowledge work, emphasizing progressive discovery and programmatic tool calling. Speakers also explored multi-agent orchestration, with AgentCraft demonstrating a gaming-inspired interface for managing agent teams, and discussions on the challenges of agent-generated code quality, the need for human judgment, and the concept of "friction" in development. The conference also covered the impact of faster AI models like Codex Spark on developer workflows, the importance of context management, and the use of AI for internal tools and conference management.
Key takeaway
For AI Engineers and ML Architects navigating the accelerating pace of AI development, prioritize integrating human judgment and strategic friction into your workflows. Focus on designing modular, agent-legible codebases and employ multi-agent systems with clear validation contracts to manage complexity and maintain quality. Your role shifts from pure code generation to orchestrating AI capabilities, demanding a deeper understanding of model strengths and weaknesses to ensure robust, high-quality software delivery.
Key insights
AI engineering is rapidly evolving, demanding new approaches to agent orchestration, code quality, and human-AI collaboration.
Principles
- Prioritize human judgment and intentional friction in AI-driven development.
- Design agent systems with modularity and clear validation contracts.
- Leverage diverse AI models for specialized tasks to optimize performance.
Method
Implement multi-agent systems with distinct roles (orchestrator, workers, validators) and structured handoffs. Utilize external memory systems like a four-file system (agents.md, plan.md, progress.md, verify.md) for persistent context and use CLI tools for agent-friendly eval management.
In practice
- Use faster models like Codex Spark for execution and larger models for planning.
- Bake validation (testing, linting, diff reviews) into every step of the workflow.
- Employ tools like AgentCraft or Command and Control for remote agent management.
Topics
- Gemma 4
- Monocontext Protocol
- Coding Agents
- Agent Orchestration
- Software Quality
Best for: AI Engineer, Machine Learning Engineer, AI Architect
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.