The AI agent revolution feels stuck in first gear. For all the excitement, only a few domains like coding, go-to-market, and search have seen products that work reliably and have achieved meaningful adoption at scale. But innovation has been bubbling beneath the surface — innovation that will unlock AI agents across a much broader set of domains.
Some feel that widespread adoption is blocked by the limitations of current AI reasoning capabilities — that we need to wait for the next generation of more sophisticated models in order to build more sophisticated agents. That’s wrong. It misses what clever practitioners are already demonstrating: Through innovative software architecture and system design, we can achieve far better reasoning and reliability from existing models than previously thought possible. Even if AI reasoning capabilities were to stand still (which they won’t), there are years of innovation ahead in how we architect, compose, and deploy these capabilities.
The missing pieces — and, therefore, the opportunities for founders — lie in three critical areas. First is imagination: Who are the customers, what are the use cases, and how do we create user experiences that make AI Agents feel natural and intuitive? Second is user trust: teams won’t adopt agents in critical workflows until they prove themselves in lower-risk tasks. The third factor is infrastructure: How do we build the agent stack with all its necessary components — memory, persistence, authentication, orchestration, and more — to make these applications reliable and production-ready?
Beyond Raw Models: The Rise of AI Software Architecture
The past year’s fixation on model capabilities has obscured a crucial insight: Through clever software architecture, we can achieve sophisticated reasoning capabilities from today’s models. Smart founders are realizing they can build significant businesses not by waiting for better models, but by making existing ones more capable through system design.
Think of foundation models as powerful but raw reasoning engines. What’s missing is the corresponding software stack that turns these engines into practical machines. Entire categories of infrastructure are emerging: tool usage to extend agents’ capabilities, auth layers that provide security, orchestration systems that coordinate graphs of multiple sub-agents, and interfaces that make these systems accessible and useful.
These architectural innovations enable capabilities that seemed out of reach. Some systems implement a “thinking fast and slow” approach, where quick responses are validated by a more deliberative reasoning layer. Others use voting mechanisms where multiple agents collaborate and cross-check each other’s work. We’re seeing verification systems that can catch errors, orchestration layers that manage complex workflows, and memory systems that maintain context across interactions.
The scope extends beyond pure architecture. Agents need to interact with real-world systems — from searching the web and accessing databases to communicating through enterprise channels and even executing payment transactions. When an agent makes a payment, posts an update to Slack, or queries a database for analysis, it needs to do so on behalf of its user with the least necessary permissions required to accomplish that task. They also need to interact via audio and video.
A rich ecosystem of tooling is emerging to support this infrastructure layer. Frameworks like Meta’s Llama Stack, OpenAI’s Swarm, Microsoft’s Magnetic One, and Amazon’s Multi-Agent Orchestrator provide industrial-strength foundations, while open-source tools like LangChain, LlamaIndex, AutoGen, BAML, Letta, and Griptape are accelerating development. Extending these capabilities — Stripe has released a payments SDK for AI Agents that works with Vercel, LangChain, and Crew AI. But we’re still in the early stages, with massive opportunities for founders to build critical pieces of this stack.
Where to Build: Four Agent Patterns That Matter
The most exciting opportunities for founders lie in reimagining how AI agents can create value across different domains. Here are four emerging patterns that suggest where to focus:
Next-Gen Copilots
Copilots are already familiar in domains like coding, but they’re evolving beyond simple completion tools. Modern copilots can understand context across multiple conversations, maintain awareness of broader goals, and proactively suggest approaches. The opportunity isn’t just in building better copilots, but in identifying new domains where this pattern can transform work. Some examples of next-gen copilots are Bolt.new, AirOps, Colimit, and even Perplexity (including its new shopping feature).
Teammate Agents
More ambitious are “teammate” agents that handle complex, multi-step tasks. These might digest hundreds of customer interviews to surface patterns, continuously monitor competitors’ products and pricing to spot strategic shifts, or analyze product usage data to identify improvement opportunities. The key is that they maintain context and chain together multiple steps – they don’t just respond to prompts, they drive workflows forward. Some examples of teammate agents include Ravenna, Sailplane, Cleric, Basepilot, and Factory.
Agent Organizations
Next are agent organizations – systems where multiple specialized agents collaborate and interact. We’re seeing fascinating examples like “AI towns” where agents interact in simulated environments, and Aaru, where a crowd of self-organizing agents run probabilistic simulations to make predictions. These point to entirely new ways of solving complex problems through emergent agent behaviors.
Invisible Agents
Finally, some of the biggest opportunities lie in making agents disappear into familiar tools, becoming developer-facing rather than user-facing abstractions. Developers will build applications on top of agents, but the end products won’t necessarily look “agentic” to users – think spreadsheets that automatically suggest analyses, development environments that write and test code alongside you, or communication tools that draft responses in your voice. The winners here will be those who make AI feel like a natural extension of existing workflows rather than a bolt-on addition. Paradigm, Perhaps, and Lutra are examples of products that use agents under the covers but present novel UX’s. Stagehand from Browserbase, Reworkd, and Tinyfish are examples of the kinds of “invisible agents” that developers can leverage to build more abstracted experiences.
Putting Agents to Work in 2025
For founders, here’s the good news:
First, as the infrastructure and software patterns we’ve discussed take shape, the map of agent opportunities is expanding dramatically. From next-gen copilots to teammate agents to agent organizations, and from user-facing applications to developer infrastructure, founders can pursue what they uniquely understand and care about. The field is too broad for any one approach to dominate — the best opportunities will come from deeply understanding a specific domain or piece of the stack.
Building user trust will be key to unlocking these opportunities. Teams are unlikely to adopt agents for mission-critical workflows until they’ve proven their value in lower-risk tasks. Starting small and iterating in trustworthy ways can pave the path to broader adoption.
Second, the other missing ingredients for success — imagination about what’s possible and deep systems engineering — are exactly what today’s technical founders have in abundance. You don’t need to compete with AI labs on model development. You just need to apply your creativity and engineering skills to applying these powerful new primitives.
So let’s get to work. The agent revolution is about to shift into high gear.
Madrona is actively investing in AI agents, the apps that rely on them, and the infrastructure that supports them. We have backed multiple companies in this area and will continue to do so. We would love to meet you if you’re building AI agents, applications, or infrastructure. You can reach out to us directly at: [email protected]