AI agents are moving from experiments to infrastructure. A year ago, most teams were wiring prompts together and calling it agentic. Today we are seeing something more structured: managed AI agents, hosted and orchestrated systems that can reason, use tools, access APIs, and run workflows without building the full stack from scratch.
The landscape
Current managed AI agent options span both hosted and open ecosystems. Platforms like Anthropic and emerging enterprise offerings from major AI providers focus on hosted agents with built-in orchestration, tool use, observability, and deployment infrastructure, reducing operational complexity for teams. Open Agents takes a more open and portable approach, enabling developers to run and control their own agents and infrastructure. The space is still evolving quickly, but the common trend is clear: AI agents are becoming managed infrastructure rather than isolated prompts or scripts.
Managed vs self-hosted
Fully managed platforms trade control for speed. Hosted agents ship with orchestration, tool use, and observability baked in, so teams spend less time on runtime plumbing.
Pros:
- Faster development
- Less infrastructure to maintain
- Better scaling and monitoring out of the box
- Strong model integration and tooling
Cons:
- Vendor dependency
- Less control over runtime and orchestration
- Data and privacy considerations
- Potential cost growth at scale
Open and self-hosted approaches, including Open Agents, follow a different philosophy: bring your own models, infrastructure, and workflows. Instead of consuming an agent as a service, you own the stack.
Pros:
- Full control
- Infrastructure portability
- Greater customization
- Better alignment with privacy-sensitive or domain-specific workflows
Cons:
- More engineering effort
- Operational complexity
- Responsibility for monitoring, scaling, and reliability
The hybrid middle
Frameworks and platforms such as LangGraph Platform and CrewAI are gaining traction for teams that want to orchestrate long-running or multi-agent workflows while retaining flexibility over models and deployment choices. LangGraph builds on the LangChain ecosystem; CrewAI focuses on role-based multi-agent coordination.
Many teams combine that layer with managed models: orchestrate agents themselves through frameworks, internal APIs, and custom logic while renting inference from a provider. That pattern tends to work when product differentiation lives in workflow, UX, or domain knowledge rather than model hosting.
When to run agents privately
Private or self-hosted agents usually make sense when the workflow touches sensitive data, regulated domains, proprietary reasoning, or business-critical automation. Legal systems, financial tooling, internal copilots, and enterprise knowledge workflows often need that model, especially when trust and data ownership matter. Institutional constraints show up here more than in a sandbox demo.
What changed
There is no universal winner. Managed agents optimize for speed. Private agents optimize for control. Most products will live somewhere in between.
The shift is not whether agents are managed or self-hosted. Agents are becoming a new layer of software infrastructure. Much like cloud platforms abstracted servers, managed agents are beginning to abstract intelligence. The real work now is deciding where that intelligence should live and how much of the loop you own.
That decision connects directly to how we already build with agents in production. In Engineering in the AI Era, I described agents as an LLM in a loop with tools: editor, tests, browser, terminal, deploy pipeline. Managed platforms productize that loop. Self-hosted stacks keep every step visible. Evolution of AI UX argued the bottleneck moved from model quality to what the system lets you do with it. Managed agents are one answer to that bottleneck: less plumbing, more execution surface.
For tool wiring specifically, standards like Model Context Protocol matter whether the agent runs on a vendor host or yours. The interface between model and tools is where portability still lives.
Managed AI agents are not a replacement for engineering judgment. They are a deployment choice about speed, control, and operational surface area. Teams that treat them as infrastructure, not magic, will get the most out of either path.