📡 Daily AI Research Briefing — May 25, 2026

Top 5 stories from the AI landscape: the agent pivot accelerates, infrastructure matures, and economics reshape the competitive map.

1 All Model Labs Are Now Agent Labs

The biggest story of the week crystallized in a single Greg Brockman quote ahead of OpenAI's likely IPO filing: "the model alone is no longer the product."

This isn't just OpenAI. It's a sector-wide pivot:

• AI21 shuttered their model team to focus entirely on agents
• DeepSeek is building their first "Harness team"
• OpenAI is shipping Codex CLI improvements weekly

The strategic implication is clear: when frontier models converge in capability, the value moves to the systems you build around them. The winning product is now model + harness + workflow + UI + memory + economics.

2 MCP Protocol Goes Stateless

The Model Context Protocol's 2026-07-28 release candidate shipped with a fundamental architectural change: statelessness.

Key changes:

• No handshake required
• No session ID management
• Any request can hit any server instance

For infrastructure teams, this is a massive operational simplification. Easier scaling, simpler load balancing, no sticky-session concerns. The protocol is maturing for production agent deployments at scale.

3 DeepSeek's Permanent 75% Price Cut Reshapes Economics

DeepSeek made their V4-Pro discount permanent this week, triggering strong reactions across the ecosystem. The numbers are stark:

As one researcher put it: we're approaching "intelligence too cheap to meter." This changes the unit economics for every AI product built on frontier models.

4 GitHub Trending: The Agent Infrastructure Boom

This week's trending repositories tell a clear story. The hot projects aren't models — they're infrastructure for making models useful:

Also trending: anthropics/knowledge-work-plugins (14K stars) for Claude Cowork, and rohitg00/ai-engineering-from-scratch (17K stars) for learning agent engineering.

5 Agent Compilation Breakthrough

A research paper highlighted this week shows that full agentic workflows — multi-step calls, tool use, scratchpads, decision structure — can be distilled into model weights and run at ~100x lower inference cost while preserving near-frontier quality.

This is one of the clearest technical arguments yet for "compiling" expensive runtime agent loops into cheaper deployable models. The economic implications are significant: what requires a full agent harness today might run as a single model inference tomorrow.

Quick Hits

• Gemini Spark — Google's 24/7 personal AI agent for recurring tasks and workflows launched at I/O
• CoreWeave Sandboxes — Public preview launched for RL, agent tool use, and model eval
• Chrome DevTools MCP — 41K stars for browser automation via Model Context Protocol
• Qwen3.7-Max — Alibaba's latest showing strong instruction following and context reliability

The Pattern

Five stories, one pattern: the agent infrastructure wave is maturing.

The labs that built foundation models are now racing to build the systems that wrap them. The open-source ecosystem is filling in the gaps — memory, knowledge graphs, sandboxes, protocols. And the economics are shifting fast enough to rewrite product strategies overnight.

If you're building AI products, the message is clear: don't bet on model lock-in. Bet on harness sophistication. The model is a commodity. The system you build around it is the product.

---

Sources: Latent Space, GitHub Trending, X/Twitter, Hacker News. Compiled by Adil Islam for Prompt Engines Lab Notes. ← All articles