The model wars dominate AI coverage. Opus 5 launched last Friday, GPT-6 arrived earlier this month, and DeepSeek went public in Shanghai last week. The infrastructure beneath these models is shifting just as fast. Four changes are crystallizing simultaneously in how models connect to tools, retrieve information, produce structured data, and orchestrate work. Each one overwrites assumptions that most teams adopted fewer than two years ago, and each introduces costs of its own.
The protocols have solidified
Anthropic open-sourced the Model Context Protocol in November 2024 as a standard way for AI models to connect to external tools and data sources. The protocol collapses what had been an N×M integration problem, in which every AI application needed a custom connector for every tool, to N+M. By mid-2026, its SDKs saw roughly 97 million monthly downloads, and more than 101,000 servers were indexed across its registries. Claude, ChatGPT, Gemini, and Microsoft Copilot all provide native support. Anthropic, OpenAI, Google, Microsoft, and AWS co-govern the protocol under the Linux Foundation’s Agentic AI Foundation. Pinterest’s production deployment runs roughly 66,000 monthly tool invocations and saves an estimated 7,000 hours per month. The ecosystem’s growth outpaced its security. AgentSeal found flaws in 66% of scanned MCP servers, and more than thirty CVEs were filed in the first two months of 2026.
OpenAI, Anthropic, and Google now guarantee that model outputs will conform to supplied JSON schemas, using constrained decoding that adds under 40 microseconds of overhead per token. Parse failure rates dropped from a range of 3–10% to zero for supported features. This guarantee shifted the failure surface. Malformed JSON used to crash a pipeline visibly. Schema-valid JSON with hallucinated field values flows silently into downstream systems. The developer’s job moved from parsing to semantic validation of every field.
Juggling context
RAG framework usage grew 400% between 2024 and 2026, and 60% of production LLM applications still use retrieval. The architecture of these systems is changing even as adoption accelerates. The standard 2024 approach pushed pre-retrieved context into prompts before the model saw a query. The emerging pattern lets agents pull data at runtime through tool calls, fetching information as they determine that they need it. VentureBeat’s Q1 2026 data showed in-house retrieval stacks growing from 24.1% to 35.6% of enterprise deployments and hybrid retrieval adoption intent tripling from 10.3% to 33.3% between January and March.
The cost gap explains retrieval’s persistence. RAG averages $0.00008 per query, while long-context approaches average $0.10, roughly 1,250 times more expensive. Long-context models still suffer from the lost-in-the-middle problem, losing 30% or more of their accuracy when relevant content sits deep inside a window. The successor combines vector search for narrowing large corpora with long-context models for reasoning over the results, absorbing retrieval into a broader context architecture.
The costs of coordination
Multi-agent orchestration produced the widest gap between expectation and outcome. In MIT testing, adding relay stages without new exogenous signals degraded accuracy from 90.7% to 22.5%, below chance baseline. Google’s 2026 scaling tests showed that multi-agent variants reduced sequential planning performance by 39–70%. A cascade study demonstrated that a single false statement achieved 100% system-wide infection across MetaGPT, LangGraph, CrewAI, and AutoGen in the absence of governance layers.
Three topologies survived: sequential pipelines, hub-and-spoke orchestration, and heavily bounded peer collaboration with phase gates. Hub-and-spoke emerged as the dominant production pattern. Google’s results showed that centralized architectures amplified errors by a factor of 4.4, while independent agents amplified them by 17.2. Anthropic’s research agent, which spawns three to five subagents in parallel, achieved a 90.2% performance gain over single-agent Opus 4 and consumed roughly fifteen times the tokens. Meta’s Ranking Engineer Agent doubled model accuracy while reducing staffing from two engineers per model to three engineers across eight. Both Anthropic and Shopify recommend starting with a single strong agent and adding agents only when that agent’s capabilities fall short.
The standards are settling
The integration layer, the retrieval pattern, the output format, and the orchestration architecture all shifted within the same twelve-month window. Teams built the 2024 versions of these layers for rapid experimentation, using custom connectors, static retrieval pipelines, defensive JSON parsing, and multi-agent architectures. Emerging standards have replaced each of these assumptions, and those standards carry their own costs, from MCP’s server-side vulnerabilities to structured outputs’ invisible semantic errors to multi-agent orchestration’s fifteen-fold token overhead. The infrastructure stack is settling around new defaults.