The AI agent space has grown dramatically over the past two years. Big tech companies are embedding agent capabilities into their existing products while a wave of startups is building purpose-built systems for specific industries and workflows.
Here is a look at the companies doing the most interesting and impactful work.
Big Tech: Platforms at Scale
OpenAI remains the most influential player. GPT-4 and GPT-4o power a large share of the agents on the market, and their Assistants API with tool use, code interpreter, and retrieval is one of the most widely used foundations for building agents. Their move into operator-level agents with products like Operator marks a direct push into consumer and enterprise automation.
Anthropic is increasingly agent-focused. Claude 3 and Claude 3.5 score consistently well on reasoning benchmarks and their computer use capability, which lets agents control desktop applications, represents a significant step toward general-purpose automation. Many developers prefer Claude for agents that need reliable, careful reasoning.
Google DeepMind has invested heavily in agents through Gemini. The Gemini function calling API is well-integrated with Google Workspace, giving agents native access to Gmail, Drive, and Calendar. Project Mariner, their browser agent, and Jules, their coding agent, show where they are heading.
Microsoft has embedded Copilot agents across the entire Microsoft 365 suite. For enterprises already on Microsoft infrastructure, Copilot Studio lets teams build custom agents without deep technical expertise.
Frontier Startups
Cognition Labs built Devin, the first AI software engineer to receive significant real-world validation. Devin can complete full software tasks end to end and is being tested by engineering teams at companies across the industry. Cognition raised at a valuation that reflects how seriously the market is taking autonomous coding.
Cohere focuses on enterprise AI with strong retrieval and tool use capabilities. Their platform is popular with companies that need on-premise or private cloud deployments rather than sending data to OpenAI or Anthropic.
Adept AI has been building agents that can operate software on behalf of users, targeting enterprise workflows that involve legacy systems and complex GUIs. Their approach is less about language and more about direct system interaction.
Dust is a well-regarded platform for building internal AI agents connected to company knowledge bases. Teams use it to build agents that can search documentation, answer questions from Slack, and automate internal processes.
Open Source Leaders
The open source ecosystem for AI agents is vibrant. LangChain remains the most widely adopted framework for building agent systems, with a large community and extensive integrations. AutoGPT was one of the earliest demonstrations of autonomous agents and spawned a wave of follow-on work.
CrewAI has emerged as a popular framework for multi-agent systems, where teams of specialized agents collaborate on complex tasks. It is well-suited for research and analysis workflows.
OpenHands (previously OpenDevin) is one of the strongest open-source coding agents, with active development and solid benchmark performance.
Infrastructure and Developer Tools
Several companies are not building agents but are building the infrastructure that makes agents possible. LangSmith (from LangChain) helps teams debug and monitor agent behavior in production. Helicone and Braintrust provide observability for LLM applications. Modal and E2B provide sandboxed code execution environments that agents need to run code safely.
This infrastructure layer is increasingly important as teams move from demos to production deployments where reliability and cost visibility matter.
What to Watch
The most interesting frontier right now is multi-agent systems where specialized agents hand off work to each other. We are also seeing more vertical-specific agents built for industries like healthcare, legal, and finance where general-purpose tools are not precise enough.
The companies that figure out reliable long-horizon task execution (agents that can work for hours without going off the rails) will define the next phase of the market.
To browse agents from many of these companies and compare what is available, AgentFilter has an updated directory covering tools across every category.