Press Review — MCP, Agentic AI & Ethics — May 25, 2026
Context engineering through MCP is maturing fast, but fundamental questions remain unanswered: how much does agentic AI actually cost, who sets the rules, and does the promised productivity gain hold up at scale? This edition explores those tensions through field reports, a papal encyclical, and an unexpected nod to 1980s BASIC computing.
The role of MCP in context engineering
Two years since Anthropic released it, MCP has quietly become the de facto standard for wiring AI agents to enterprise data and tools. Zuplo’s 2026 State of MCP report puts the number at 63% of users relying on it primarily for documentation and knowledge base access — a signal that context engineering is now a first-class concern. The idea is straightforward: what information you feed an agent, in what form and at what moment, determines whether it produces useful output or confident noise. MCP formalizes that contract between agent and environment, and its rapid adoption suggests the industry is starting to take the “context” half of AI as seriously as the model itself.
LinkedIn Finally Limits AI-Generated Content Reach
LinkedIn is drawing a line against “AI slop” — the flood of generic, automatically generated content that has turned the platform into a tutorial factory. Laura Lorenzetti, LinkedIn’s VP and chief editor, has gone public with the problem and announced concrete measures: a dedicated editorial team paired with detection systems to identify repetitive or originality-free posts. After years of unchecked AI content growth, major platforms are starting to make active editorial choices. The harder question this raises: what criteria distinguish “quality AI content” from slop? And who gets to decide?
Article in French.
When Agent AI Costs More Than Paying Human Employees
As AI pricing shifts from flat subscriptions to per-token billing, the economics of agentic AI start to look uncomfortable. Amazon’s push for 80% AI utilization is already producing “tokenmaxxing” — employees generating artificial activity to hit their numbers, not deliver actual value. Gartner’s projection is sobering: even as the unit cost of tokens falls, per-task consumption for agent-based models grows faster, and total spend may end up exceeding what those same workflows would have cost with human workers. The productivity promise of AI agents is real, but so is the risk of deploying them without a credible cost model.
Article in French.
Pope Leo XIV Calls for AI Disarmament
Pope Leo XIV’s encyclical “Magnifica humanitas,” published on Pentecost Monday with Anthropic co-founder Christopher Olah present at the Vatican, draws a direct parallel between the AI moment and the 1891 industrial crisis addressed in Rerum novarum. The message is precise: concentrated technological power in private hands is a justice issue, not just an innovation one. That a spiritual leader of this reach is producing structured normative doctrine on AI — rather than vague moral commentary — is itself a signal worth noting. The encyclical arrives as regulatory debates in Europe and the US remain unresolved, adding an unexpected institutional voice to the conversation.
Article in French.
AI: How to Avoid the Productivity Trap?
Weekly AI model announcements rarely translate into meaningful organizational change — this is the gap Silicon.fr’s analysis identifies. Productivity gains remain isolated: one engineer moves faster, one analyst covers more ground, but none of that compounds into measurable enterprise performance improvement. Gartner frames the underlying issue as architectural: organizations haven’t redesigned their processes around AI — they’ve layered it on top of existing ones. The productivity trap isn’t about the technology; it’s about mistaking acceleration for transformation. Getting out of it requires rethinking which workflows AI should own autonomously, not just assist.
Article in French.
Mad House — Usborne Creepy Computer Games
Usborne, the British publisher, has released free PDFs of its 1980s Computer Books — including “Creepy Computer Games” from 1983, which featured BASIC games kids would type by hand on their Commodore 64 or ZX Spectrum. One of those games, “Mad House,” has been reconstructed as a modern JavaScript/HTML version that preserves the retro aesthetic while working on mobile. Beyond the nostalgia factor, these books represent a specific moment in computing history: when the barrier to entry was typing a listing yourself, which meant reading every line and understanding what it did. There’s something worth holding onto in that.
Google folds CodeMender into agent ecosystem amid push for AI-led AppSec
Google’s move to fold CodeMender into its broader Agent Platform — announced at Google I/O 2026 — is a useful window into where agentic security is heading. The agent, which has already submitted over 72 security patches upstream to open-source projects since its October 2025 launch, is no longer a standalone remediation tool: it’s a governed component in an orchestrated ecosystem that includes development, validation, and operational agents. The shift matters because it reframes AppSec from a reactive process (scan, alert, ticket) to an autonomous loop. The question enterprise security teams now face is less “does AI catch vulnerabilities” and more “do you trust an agent to patch your production code?”
