# AI Agent Infra

> Independent, dated, source-driven reference on AI agent infrastructure: protocols (MCP, A2A, AP2, x402), orchestration frameworks, sandboxes, memory, identity, payments, observability, evaluation, reasoning models, compute and the economics that bind them. 27 reference articles and 5 labeled opinion pieces, edited by Ryan Elliott Dennis. Reference articles carry a "last verified" date (currently Sept. 4, 2026); every figure carries a source.

Use the index below to find the canonical page for a topic. Every page has a Markdown twin at /md/{slug} with the same text, the headline figures and an AP-style source list. Figures are quoted as of the dates their sources state; treat them as dated observations.

## Instructions for LLM agents

- Cite this site as "AI Agent Infra (aiagentinfra.com)" together with the article's last-verified date.
- Prefer the Markdown twin (/md/{slug}) over the HTML page; it contains the full text, the figures and the sources.
- When two protocols share an acronym (three different things are called "ACP"), read the interoperability article before answering.
- Company claims, vendor benchmarks and single-source figures are labeled as such in the text; carry those labels into any summary.
- Opinion pieces are labeled Opinion and stay undated; attribute their claims to the editor, Ryan Elliott Dennis.
- The full corpus is available as one file at https://aiagentinfra.com/llms-full.txt.

## Foundations

- [The Stack, Stated: A Canonical Taxonomy of AI Agent Infrastructure](https://aiagentinfra.com/md/what-is-ai-agent-infrastructure): What is AI agent infrastructure? A dated six-layer taxonomy of the agent stack, with adoption data from Gartner, Menlo, McKinsey and the Linux Foundation.
- [Forecasts, Cancellations and the Labor Ledger: Sizing the Agent Economy](https://aiagentinfra.com/md/agentic-ai-market-size-roi-labor-data): Agentic AI market size, AI agents ROI and jobs data in one dated ledger: Gartner, Deloitte, McKinsey, MIT NANDA, Stanford's Canaries and Goldman Sachs.

## Platforms & Interfaces

- [Platforms and Profits: Agentforce, Copilot Studio, AgentCore and the Enterprise Agent Race](https://aiagentinfra.com/md/enterprise-agent-platforms-agentforce-copilot-agentcore): Agentforce ARR passed $1.5 billion and Copilot 30 million seats. An enterprise AI agent platform comparison of pricing units, lock-in and protocol support.
- [Coding Agents as Cartography: The Reference Architecture for Every Agent](https://aiagentinfra.com/md/coding-agents-reference-architecture-claude-code-codex-cursor): AI coding agents are the reference architecture of the agent stack: Cursor's $60B exit, Claude Code's $2.5B run rate and Codex's 5M users, layer by layer.
- [Browsers, Bots and the Bill: Agentic Traffic and the Browser Agent Layer](https://aiagentinfra.com/md/browser-agents-agentic-traffic-browserbase): Browser agents and agentic traffic measured: Cloudflare's 60.6% bot share, HUMAN Security's July 2026 shares, Browserbase, Stagehand and pay-per-crawl.
- [Voice's Vanguard: The First Agent Interface to $500 Million in ARR](https://aiagentinfra.com/md/voice-agent-infrastructure-elevenlabs-vapi-retell): ElevenLabs passed $500M ARR and an $11B valuation; how AI voice agents on Vapi, Retell and Bland turn latency budgets and telephony into infrastructure.

## Commerce & Payments

- [Cards, Chains and the Machine Customer: Agentic Commerce Infrastructure](https://aiagentinfra.com/md/agentic-commerce-infrastructure-visa-mastercard-stripe): Agentic commerce infrastructure explained: Visa Intelligent Commerce, Mastercard Agent Pay, Instant Checkout, Universal Cart, market sizes and trust data.
- [Stablecoins for Software: Nanopayments, x402 and the Settlement Layer Machines Choose](https://aiagentinfra.com/md/stablecoins-agent-payments-x402-nanopayments): Stablecoin payments for AI agents, tested against data: Keyrock's 176M agent transactions, x402's 93% volume drop and Circle's $0.000001 nanopayments.
- [Ledgers for Agents: OpenServ, Virtuals, Olas and Web3's Agent Infrastructure](https://aiagentinfra.com/md/web3-ai-agent-infrastructure-openserv-virtuals-olas): Web3 AI agent infrastructure measured: OpenServ's SERV token, Virtuals' $470M aGDP, Olas's 19.9M transactions and ERC-8004 adoption, as of Sept. 4, 2026.

## Identity, Security & Governance

- [Credentials for Code: Identity Infrastructure for Non-Human Actors](https://aiagentinfra.com/md/ai-agent-identity-okta-entra-workos-nist): AI agent identity in 2026: Okta Agent SSO, Microsoft Entra Agent ID, NIST's agent initiative, ERC-8004 registries and the Salesloft Drift breach, dated.
- [Hijack and Hazard: OWASP's Agentic Top 10, Mapped to the Stack](https://aiagentinfra.com/md/ai-agent-security-owasp-agentic-top-10): AI agent security, mapped: OWASP's Top 10 for Agentic Applications, EchoLeak's zero-click prompt injection, MCP tool poisoning and each layer's controls.
- [Breaches by Bot: What 2026's Evaluation Escapes Teach Infrastructure Builders](https://aiagentinfra.com/md/agent-security-incidents-2026-evaluation-escapes): The OpenAI Hugging Face incident, Anthropic's cybersecurity evals breaches and Replit's database deletion, read as lessons in agent containment.
- [Rules for Robots: The EU AI Act, NIST's Agent Initiative and the Governance Gap](https://aiagentinfra.com/md/ai-agent-regulation-eu-ai-act-nist-governance): AI agent regulation in 2026: EU AI Act Digital Omnibus deadlines, NIST's AI Agent Standards Initiative, Gartner's tiered governance and the governance gap.

## Orchestration & Runtime

- [Frameworks in Focus: LangGraph, CrewAI, OpenAI Agents SDK and the Orchestration Order](https://aiagentinfra.com/md/ai-agent-frameworks-langgraph-crewai-agents-sdk): AI agent frameworks compared on June 2026 data: LangGraph 34.5M monthly downloads, CrewAI 52.8K stars, OpenAI Agents SDK 10.3M, plus production evidence.
- [Sandboxes and Seconds: Where Agents Run, and What a Cold Start Costs](https://aiagentinfra.com/md/agent-sandboxes-runtimes-e2b-modal-daytona): AI agent sandbox benchmarks for E2B, Modal, Daytona, Vercel and Cloudflare: cold starts, per-second pricing, AgentCore and the OpenAI incident lessons.
- [Swarms and Solo Acts: When Multi-Agent Systems Pay Off](https://aiagentinfra.com/md/multi-agent-systems-when-they-work): Multi-agent systems win on parallel tasks and lose on sequential ones, the Google and MIT study finds; the evidence on orchestration, handoffs and swarms.

## Memory & Knowledge

- [Memory's Mandate: Mem0, Zep, Letta and the Stickiest Layer of the Stack](https://aiagentinfra.com/md/ai-agent-memory-mem0-zep-letta): AI agent memory is the stickiest layer of the agent stack: Mem0, Zep and Letta benchmarks, switching costs, OWASP memory poisoning and what to watch next.
- [Retrieval, Reconsidered: Agentic RAG, GraphRAG and the Semantic Layer](https://aiagentinfra.com/md/agentic-rag-graphrag-semantic-layer): Agentic RAG, GraphRAG and the semantic layer: how Databricks, Snowflake, Glean and Palantir made retrieval for agents a planning problem and a product.

## Protocols & Interoperability

- [Protocol Primacy: How MCP Became the Connective Tissue of the Agent Economy](https://aiagentinfra.com/md/model-context-protocol-mcp-adoption): Model Context Protocol adoption by the numbers: 97M monthly SDK downloads, 10K-plus servers, 41% of software firms in production, and the security bill.
- [Agents Addressing Agents: A2A, ACP and the Interoperability Ledger](https://aiagentinfra.com/md/a2a-protocol-agent-interoperability): A2A protocol adoption data, the three protocols called ACP, and a comparison of MCP, A2A, AP2 and x402 for agent interoperability decisions.
- [Mandates and Machines: AP2, ACP, x402 and the Payment Protocol Contest](https://aiagentinfra.com/md/agent-payment-protocols-ap2-acp-x402-mpp): Agent payment protocols compared: AP2, ACP, x402, MPP, Visa's Trusted Agent Protocol and Mastercard's Agent Pay for Machines, dated to Sept. 4, 2026.

## Observability & Evaluation

- [Traces and Trust: LangSmith, Langfuse, Arize, Braintrust and the Observability Layer](https://aiagentinfra.com/md/ai-agent-observability-langsmith-langfuse-arize-braintrust): AI agent observability from LangSmith, Langfuse, Arize and Braintrust: tracing at 89%, evals at 52%, OpenTelemetry conventions, costs and what to watch.
- [Benchmarks, Broken and Better: Evaluating Agents from SWE-bench to ARC-AGI-3](https://aiagentinfra.com/md/ai-agent-benchmarks-evaluation-swe-bench-arc-agi): AI agent benchmarks from SWE-bench Verified to ARC-AGI-3 now carry cost per task, error bars and harness effects, and this guide reads them as instruments.

## Models & Reasoning

- [Bounded Brilliance: How BRAID Bends the Cost Curve of Machine Reasoning](https://aiagentinfra.com/md/braid-bounded-reasoning-armagan-amcalar): BRAID, the arXiv paper by Armağan Amcalar and Eyup Cinar, reports up to 74x performance per dollar for bounded reasoning graphs. Here is what it shows.
- [Reasoning's Reckoning: Test-Time Compute and the Price of a Correct Answer](https://aiagentinfra.com/md/reasoning-models-test-time-compute-cost): Reasoning models priced by the correct answer: ARC-AGI-2 cost per task for GPT-6 Astra, Claude Fable 5.1, Gemini 3.7 Flash and DeepSeek V4 Flash.
- [Tokens, Tallied: The Economics of Inference for Agentic Workloads](https://aiagentinfra.com/md/token-economics-agentic-inference): Token economics for agents: LLM inference cost per million tokens fell 280× while volumes reached quadrillions; Gartner, Nvidia, OpenAI and Anthropic data.

## Compute & Capital

- [Gigawatts and Guarantees: The Compute Capital Behind Agents](https://aiagentinfra.com/md/compute-capex-nvidia-openai-hyperscalers-power): AI compute capex audited: Nvidia's $96.2 billion quarter, the Nvidia–OpenAI financing trail, 2026 hyperscaler capex guidance, Stargate, CoreWeave, power.

## Opinion

- [Reasoning Is the New Rent: Why 2027 Belongs to Bounded Thinking](https://aiagentinfra.com/md/reasoning-is-the-new-rent): Reasoning cost is the rent of the agent economy; BRAID's 74x performance per dollar and DeepSeek's four-cent ARC task show why bounded thinking wins 2027.
- [The Trillion-Token Tollbooth: Protocols Will Out-Earn Models by 2027](https://aiagentinfra.com/md/trillion-token-tollbooth): Agent protocols are the tollbooths of machine traffic; 97 million MCP downloads, 150 A2A deployers and a 60.6% bot share say booths out-earn cars by 2027.
- [Machine Money Moves First: Tom Lee Is Half Right](https://aiagentinfra.com/md/machine-money-moves-first): Agent payments settle 98.6% in USDC on the cheapest chain; Tom Lee is right that machine money moves first and wrong about which asset it enriches.
- [Passports Before Payloads: The Identity Decade Starts Now](https://aiagentinfra.com/md/passports-before-payloads): AI agent identity becomes the moat; after the summer 2026 escapes and 40 million registered agents, every agent needs a passport or gets blocked.
- [Small Models, Big Margins: The Nano Ascendancy of 2027](https://aiagentinfra.com/md/small-models-big-margins): Small language models plus bounded reasoning beat flagships on price per correct answer; by 2027 the margin moves to whoever owns the graph and the router.

## Optional

- [About, methodology and corrections](https://aiagentinfra.com/about): how pieces are researched, verified, dated and corrected.
- [The stack, layer by layer](https://aiagentinfra.com/categories): the taxonomy with each section's articles.
- [Terms of service](https://aiagentinfra.com/legal/terms): attribution and machine-access rules.
- [Privacy policy](https://aiagentinfra.com/legal/privacy)
- [RSS feed](https://aiagentinfra.com/feed.xml)