Takeaways & Lessons from AI Engineer Paris 2025 (alexspyropoulos.com)

At AI Engineer Paris 2025 the dominant themes were practical hardening of agents and a budding infrastructure shakeup. Observability for agentic systems emerged as table stakes: vendors (think Sentry-style debugging for agents) are building dashboards, error traces, and execution metrics, but fragmentation is a real headache—expect a “Mixpanel for Agent Observability” and a “Segment for Agent data” to appear soon. Parallel to that, the “DIY cloud” trend is accelerating: Firecracker-style microVMs are enabling specialist providers to offer Lambda-like serverless on bare metal, carving out niche compute offerings that will nibble at big-cloud revenue. The community has shifted focus after GPT-5 disappointment back to engineering fundamentals: unit testing, compliance testing, maintainability and purpose-built agents (e.g., Datadog-style agents that take action on alerts). Practically useful tools and concepts flagged at the event include MicroVM/Firecracker, Docker Hub MCP Server, roocode (open-source coding agent), cagent (Dockerfile-style agent builder), daytona.io (secure elastic infra), and Light LLM (inference/serving). Important conceptual takeaways: “thinking traces” (LLM reasoning steps exposed via OpenAI’s Responses API), “context rot” (performance drop over long contexts), the “needle-in-a-haystack” retrieval benchmark, and speech diarization (speaker segmentation). Together these trends imply a near-term boom in agent-native analytics, standards for observability pipelines, creator-monétized MCP servers, and a more heterogeneous compute layer driven by microVMs and specialized infra.