AI Agents & Agentic Workflows

The Landscape Right Now

We’re at an inflection point where the hype is catching up to reality — but the gap between demos and production is where all the money will be made.

The AI agents market sits at roughly $7-8 billion in 2025, with every analyst projecting 40-50% CAGR through the end of the decade. That’s the kind of growth curve that created AWS, Stripe, and Shopify. But here’s what matters more than the top-line number: fewer than one in four organizations that are experimenting with agents have successfully scaled them to production. That gap — between “cool demo” and “reliable system” — is 2026’s central business opportunity.

The landscape has three tiers forming right now:

Tier 1: The hyperscalers — Microsoft, Google, Amazon, and the model providers (OpenAI, Anthropic) are building the foundational infrastructure. They’re competing to be the “cloud” of the agentic era. Every major AI lab now ships its own agent framework: OpenAI has the Agents SDK, Google released ADK, Anthropic shipped the Agent SDK, Microsoft has Semantic Kernel and AutoGen. This consolidation signals where these companies believe the value will concentrate.

Tier 2: Enterprise platform vendors — Salesforce (Agentforce), ServiceNow, UiPath, and others are embedding agents into existing enterprise software. They’re betting that agents are a feature, not a product. They’re right for their customers and wrong about the broader market.

Tier 3: Agent-native startups — This is where the disruption happens. Companies building products where autonomous agents are the primary interface, not a bolt-on. Sierra for customer service, Cognition (Devin) for coding, and hundreds of vertical-specific startups. Agent startups raised $3.8 billion in 2024, nearly tripling the previous year.

On the infrastructure side, four protocols have emerged as the communication stack: MCP (Anthropic’s Model Context Protocol, 97 million SDK downloads), A2A (Google’s Agent-to-Agent protocol, 50+ partners), ACP, and UCP. These aren’t competing — they’re layers. MCP handles tool access, A2A handles agent coordination, and ACP/UCP handle commercial transactions. The Linux Foundation launched the Agentic AI Foundation (AAIF) in December 2025, co-founded by OpenAI, Anthropic, Google, Microsoft, AWS, and Block, to govern these standards. That’s a strong signal that the infrastructure is stabilizing.

The framework landscape is crowded but clarifying: LangGraph leads for complex Python multi-agent orchestration, CrewAI for rapid role-based prototyping, and Mastra for TypeScript teams. No-code builders like n8n (150k+ GitHub stars) are becoming the action layer for non-technical builders. Over 120 production-ready tools now exist across 11 categories.

Developer pain is real and well-documented. A study of 3,191 Stack Overflow posts found installation and dependency conflicts top the list at 21%, followed by orchestration challenges at 13% and RAG engineering at 10%. The hardest problems — tool coordination, observability, and cost management — persist despite all the framework competition. Agents behave differently under load. An orchestration pattern working at 100 requests/minute can collapse at 10,000. Most teams can’t see nearly enough of what their agentic systems are doing in production.

The economics are brutal if you’re not careful. Each agent action involves one or more LLM calls. When agents chain dozens of steps per request, a workflow costing $0.15 per execution becomes terrifying at 500,000 daily requests. Cost optimization is becoming an architectural concern on par with cloud cost management.

Where This Is Going

Thesis: The next 24 months will see the “microservices moment” for AI agents — monolithic all-purpose agents give way to orchestrated teams of specialized agents, and the real money moves from building agents to building the infrastructure that makes agents reliable, observable, and affordable.

Three shifts matter most:

1. From Single Agents to Multi-Agent Systems

Gartner reported a 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025. That’s not hype — it’s enterprise architects realizing that one giant agent trying to do everything fails the same way monolithic applications fail. The winning pattern is “puppeteer” orchestrators coordinating specialist agents, each scoped to a narrow domain. This creates an enormous need for coordination infrastructure, observability, and governance tooling that barely exists today.

2. The SaaS Model Is Breaking

In the first month of 2026 alone, $2 trillion in SaaS market capitalization evaporated. When one AI agent can replace dozens of human software licenses, per-seat pricing collapses. This isn’t a market correction — it’s structural disruption. The companies that survive will shift to outcome-based or usage-based pricing. The ones that thrive will be agent-native from day one, never having carried the baggage of per-seat economics.

3. The “Creator Era” of Software

Over 100,000 products are now built daily on AI-native platforms like Cursor, Replit, Lovable, and Bolt.new. Cursor went from zero to $1 billion ARR in 24 months. The bottleneck is shifting from “can we build this?” to “should we build this, and for whom?” Non-technical builders — plumbers, parents, domain experts — are entering the ecosystem because the abstraction layer fundamentally changed. This creates a massive new market of people who need agent-powered tools but don’t think in terms of “agents.”

What’s Coming in 12-24 Months

• Agent-to-agent commerce: Agents will negotiate, transact, and settle payments with each other. The ACP and UCP protocols are early versions of this. First movers in agent payment infrastructure will build durable businesses.
• Agent observability as a category: Just as Datadog emerged for cloud monitoring, we’ll see dedicated platforms for tracing, debugging, and optimizing multi-step agent workflows. Traditional monitoring doesn’t cut it when the same input produces different execution paths.
• Vertical agent platforms: The horizontal frameworks (LangGraph, CrewAI) are settling. The next wave is verticalized agent platforms for specific industries — legal, healthcare, real estate, logistics — with domain-specific guardrails, compliance, and integrations baked in.
• MCP as the new REST: MCP is on track to become the universal interface between AI and the rest of the software world. Building MCP servers for underserved tools and services is a near-guaranteed business for the next 2-3 years.

The Whitespace Map

1. Agent Observability & Debugging Tools

The gap: When an agent takes a 12-step journey to answer a query, most teams have no visibility into intermediate decisions. Traditional APM tools track latency and throughput but miss the reasoning chain entirely. Debugging non-deterministic agent behavior is a nightmare.

Who feels the pain: Every engineering team running agents in production. The study from Stack Overflow data shows orchestration (13%) and evaluation (10%) are persistent, under-supported pain areas.

Why incumbents aren’t solving it: Datadog and New Relic were built for deterministic request-response patterns. Agent workflows are fundamentally different — branching, looping, tool-calling, with different paths for the same input. It requires a new mental model for observability.

How big: If agent observability follows the cloud monitoring trajectory, this is a multi-billion dollar category within 5 years. Every company running agents in production will need it.

2. Agent Cost Management (FinOps for AI Agents)

The gap: Teams are getting blindsided by agent costs. A workflow that looks cheap in testing can burn through budgets at scale. There’s no equivalent of AWS Cost Explorer for agent token spend.

Who feels the pain: Mid-size companies scaling from prototype to production. Startups who built with GPT-4-class models and now need to optimize before unit economics kill them.

Why incumbents aren’t solving it: Cloud FinOps tools don’t understand token-level economics, model routing, or the concept of routing simpler sub-tasks to cheaper models. This requires deep understanding of LLM pricing, agent architecture, and workflow optimization.

How big: If enterprises are losing 30-50% of projected AI ROI to integration overhead and cost friction (per McKinsey-style estimates), the market for tools that reclaim even a fraction of that waste is enormous.

3. MCP Server Ecosystem (The Long Tail of Integrations)

The gap: MCP has 97 million downloads and support from every major AI provider, but thousands of SaaS tools, internal systems, and niche APIs still lack MCP servers. Someone has to build the connectors.

Who feels the pain: Developers trying to give their agents access to specific business tools. Enterprise teams that need agents to interact with legacy systems.

Why incumbents aren’t solving it: The big players are building MCP servers for the top 50 integrations. The long tail of thousands of smaller tools is a perfect market for indie builders and small teams.

How big: The MCP market is expected to reach $1.8B in 2025 alone. As MCP becomes the new REST API, building and maintaining MCP servers becomes a recurring-revenue business.

4. Agent Testing & Evaluation Infrastructure

The gap: How do you test something that behaves differently every time? Agent evaluation is one of the hardest unsolved problems. Teams are shipping agents with minimal testing because the tooling doesn’t exist.

Who feels the pain: Anyone who’s had an agent hallucinate a refund policy in a demo (a real story from a developer who lost a contract over it). Every team that needs to prove their agents are reliable before deploying to production.

Why incumbents aren’t solving it: Testing frameworks are built for deterministic software. Agent testing requires new primitives: behavioral testing, trajectory evaluation, cost-per-quality tradeoffs, and regression detection across prompt changes.

How big: Every agent deployment needs testing. This is a horizontal need across the entire market.

5. “Agent-Ready” API Layer

The gap: APIs were designed for human developers — with documentation meant for humans, rate limits tuned for human usage patterns, and authentication flows requiring human interaction. AI agents consume APIs at machine speed, with parallelism, thousands of requests per second, and dynamic discovery. 51% of developers worry about unauthorized or excessive AI calls. The current API infrastructure isn’t built for this.

Who feels the pain: API providers seeing unexpected agent traffic. Developers building agents that hit rate limits, fail authentication, or can’t parse documentation programmatically.

Why incumbents aren’t solving it: Existing API management tools (Kong, Apigee) are oriented toward human developer experience. Making APIs agent-consumable requires new approaches to documentation, rate limiting, authentication, and discovery.

How big: Every API in existence eventually needs to become agent-ready. This is infrastructure for the next era of the internet.

Signals to Watch

• AAIF Events Calendar 2026: The Agentic AI Foundation has announced a global event series including AGNTCon + MCPCon in Amsterdam (Sept) and San Jose (Oct), plus MCP Dev Summits across 10 cities. Watch for which standards gain the most traction at these events — they’ll define the next 2-3 years of infrastructure decisions.

• MCP 2026 Roadmap: The protocol’s maintainers are explicitly prioritizing auth, governance maturation, and enterprise readiness. When these ship, expect a wave of enterprise adoption that’s currently held back by security concerns.

• SaaS valuation compression: If per-seat SaaS continues to lose market cap, the shift to agent-native and outcome-based pricing accelerates. Watch Salesforce, ServiceNow, and HubSpot quarterly earnings for signs of per-seat pricing erosion.

• Model cost curves: Every 2x reduction in inference cost unlocks new agent use cases that were previously too expensive. Watch for pricing announcements from Anthropic, OpenAI, and Google — cheaper models = bigger agent market.

• Regulation: The EU AI Act is being implemented and several US states are considering AI governance legislation. Agent-specific regulation (who’s liable when an agent makes a bad decision?) is coming. Companies that build governance tooling early will be positioned as the compliance layer.

• Multi-agent production deployments: Pinterest’s production MCP ecosystem (66,000 invocations/month, 7,000 hours saved/month) is an early signal. Watch for more case studies like this — they validate the architecture and create blueprints for other companies.

Contrarian Take

The biggest opportunity in AI agents isn’t building agents — it’s building the boring infrastructure that makes agents reliable.

Everyone is racing to build the flashiest agent. The venture money is flowing to “AI agents that do X.” But the actual bottleneck isn’t capability — models are already good enough for most use cases. The bottleneck is reliability, observability, cost management, testing, governance, and security. These are deeply unsexy problems. They don’t make good demos. They’ll never go viral on Twitter.

But they’re exactly the kind of problems that build $1B+ companies. AWS didn’t win by being the flashiest cloud — it won by being the most reliable. Stripe didn’t win by having the coolest payment UX — it won by handling the ugly compliance and fraud infrastructure that nobody else wanted to touch.

The indie hacker who builds the “Stripe for AI agents” — the infrastructure layer that handles the ugly, complex, essential plumbing — will build a more durable business than the one who builds the 500th customer support chatbot. The agent layer will keep churning. The infrastructure layer will compound.

The boring stuff is where the money is. It always has been.