Skip to main content

🌍 AI Engineer World Fair 2025

· 5 min read

I attended the AI Engineer World Fair in June 2025. Here are the key insights and trends I discovered.

Mind Map​

Key Technologies and Frameworks​

Model Context Protocol (MCP)​

What it is: An open standard protocol that provides a consistent interface between AI models and external tools

Key benefits:

  • Standardized communication between agents and tool servers
  • Easy extension of agent capabilities with new tools
  • Interoperability between different AI models and tool ecosystems
  • Composition of multiple tools into complex workflows

Hidden capabilities:

  • Elicitation - allows server to request completion from client
  • Dynamic discovery - expand tool set based on scope of conversation
  • Rich stateful interactions

Implementation examples:

  • Obsidian MCP server for document management
  • Neo4j MCP server for graph databases
  • Bookmark manager MCP
  • Noteplan integration potential

Nova Act​

What it is: A research preview of an SDK and model for building agents that can reliably take actions in web browsers

Key capabilities:

  • Navigate websites and execute complex web workflows
  • Break down tasks into smaller, reliable steps
  • Extract specific information from web pages
  • Run multiple browser sessions in parallel
  • Capture screenshots for debugging and documentation

Use cases: Web traversal and testing, food research, automation of web-based tasks

Strands Agents​

What it is: A simple yet powerful SDK for building and running AI agents

Key features:

  • Lightweight & flexible agent loop
  • Model agnostic - supports various providers
  • Advanced capabilities like multi-agent systems
  • Built-in MCP support

GraphRAG​

What it is: An approach to knowledge retrieval that maintains causal links in memory systems

Benefits:

  • Claims to solve hallucinations
  • Improves hypothesis generation
  • Maintains relationships between concepts

Implementation tools:

  • Neo4j for graph database
  • Congee for building custom graphs
  • Graphiti/Zep's open source graph framework

AI Agent Development​

Agent Types and Approaches​

  • Computer use agents: Agents that generate actions on screen
  • Agentic reasoning: Ability to evaluate results and update plans
  • Workflows vs. Agents:
    • Workflows: Composable pipelines with explicitly ordered steps
    • Agents: Maintain memory and are turn-based with tools
    • Agent supervisors: Agents that call other agents as tools

Agent Memory Systems​

Importance: Enables agents to maintain context and learn from past interactions

Approaches:

  • Vector databases (like Pinecone)
  • Graph-based memory (Neo4j)
  • Semantic fact generation

Challenges:

  • Semantic similarity is not business relevance
  • No one-size-fits-all memory solution

Agent Evaluation (Evals)​

Challenges:

  • Models are designed to be creative, not to judge
  • Limited standard/generic evaluation metrics
  • Guardrails fail for taste and opinion

Approaches:

  • Using LLMs to test LLMs ("vibe testing")
  • Ranking and scoring responses
  • Domain-specific evaluation metrics

Tools: WithPi.ai platform for building scoring systems

Coding with AI​

Vibe Coding​

What it is: Using AI agents for coding assistance and generation

Tools mentioned:

  • Windsurf and Cursor IDEs
  • Claude Code
  • Qodo (formerly Codium)

Best practices:

  • Start small with clear definition of done
  • Be prepared to throw code away
  • Always review AI-generated code
  • Write specs and documentation
  • Maintain a strict style guide
  • Run linters and tests

AI Coding Workflows​

  • Planning: Using AI to create development plans
  • Writing: Code generation with best practices built in
  • Testing: Automated test generation and coverage improvement
  • Reviewing: AI-assisted code review
  • Debugging: Faster issue identification and resolution

Measuring AI Coding Quality​

Metrics:

  • PCW (Percentage of Code Written)
  • Defect count
  • Time to understand new code
  • Time to write unit tests
  • Debugging speed

Emerging AI Technologies​

Thinking in AI Models​

Concept: Adding a loop to allow models to iteratively perform test-time compute

Benefits:

  • Less compute for simple requests
  • More compute for harder requests
  • Dynamic allocation of thinking resources

Implementation:

  • Trained with reinforcement learning
  • Configurable thinking budgets
  • Deep thinking for complex problems

AI Memory and Reasoning​

Concepts:

  • Brains as prediction machines
  • Hallucinations as controlled predictions
  • Reasoning requiring causal links

Approaches:

  • Traditional RAG - pulling info and enriching prompt
  • Agentic RAG - having tools for getting info
  • Deep Research - planning steps for retrieval

AI Tools People Are Using​

IDEs: Cursor, Windsurf, Claude Code Models: Claude 4 for large projects, SWE-1 for small tasks, Gemini-2.5-pro as backup Other tools:

  • v0 for rapid demo/template building
  • Replit for building test apps on mobile
  • AssemblyAI for speech-to-text
  • ElevenLabs for text-to-speech
  • Phi3:mini & deepseek via Ollama for offline development
  • Granola for organization

Notable Companies and Projects​

  • OpenRouter: LLM marketplace with single API access to multiple models
  • Featherless.ai: Personalized AGI focused on reliability
  • Upside: Forensic attribution and intelligence for revenue
  • OpenAudio: Instructurable voice model with expressive capabilities
  • Krea.ai: AI image generation platform
  • Qodo: End-to-end AI across SDLC
  • Factory: Agent-driven development platform
  • Dagger: Platform for efficient agent environments

Action Items and Resources​

Tools to Explore​

  • Amazon Q CLI and developer tools
  • Nova Act for web automation
  • Strands Agents for agent building
  • GraphRAG for knowledge management
  • MCP servers for various applications

Learning Resources​

Potential Projects​

  • Building a personal MCP server
  • Setting up a Neo4j GraphRAG system
  • Creating an agent for web traversal using Nova Act
  • Extending Noteplan with MCP integration

The AI Engineer World Fair confirmed that AI engineering is rapidly maturing from experimental territory to a standard engineering discipline. The tools are getting better, the practices are becoming standardized, and the industry is moving toward production-ready AI systems.

The future belongs to engineers who can effectively build, deploy, and maintain AI systems at scale.

📝 References

Grounded in:

  • AI Engineer World Fair 2025 conference notes (/Users/omareid/Library/Containers/co.noteplan.NotePlan3/Data/Library/Application Support/co.noteplan.NotePlan3/Notes/🏢 Amazon/🏢📝 Notes/🏢 Ai.Engineer Conference/)
  • Session recordings and presentation materials
  • Networking contacts and industry insights
  • Personal experience and key takeaways