Tech • IA • Crypto
Ship your first Managed Agent
8/10TL;DR
Anthropic introduced “managed agents,” a production-ready framework that abstracts infrastructure, scaling, and reliability, enabling developers to build and deploy AI agents significantly faster.
KEY POINTS
Shift from APIs to managed agents
Early access to Claude models relied on a basic messages API, requiring developers to manually implement core agent components such as context management, tool orchestration, and execution loops. As model capabilities improved, these responsibilities became increasingly complex. Managed agents were introduced to eliminate this burden by packaging these primitives into a unified system.
Agent SDK limitations drove new approach
The intermediate Agent SDK enabled more powerful use cases, including agents that interact with file systems via tools like Claude Code. However, developers still had to handle hosting, scaling, and safety concerns themselves. This gap led to the creation of a fully managed infrastructure layer.
Fully managed infrastructure and faster deployment
Claude Managed Agents provide built-in hosting, scaling, sandboxing, observability, and tool execution. This allows developers to focus on defining tasks, tools, and domain logic rather than infrastructure. The approach has reduced time to production by 10–15× in some cases.
Core architecture: agents, environments, sessions
The system is built around three primitives. Agents define the model, tools, and behavior. Environments act as execution containers where actions occur. Sessions connect the two, maintaining state and enabling interaction. This structure allows persistent, resumable workflows without manual state management.
Server-side agent loop and persistence
Unlike traditional setups, the agent loop runs server-side, ensuring durability even if a client disconnects. Sessions persist automatically, allowing workflows to resume without data loss. This removes the need for developers to build database layers or recovery systems.
Decoupling “brain” and “hands”
A key design decision separates reasoning (agent loop) from execution (tools and environments). This improves security by isolating credentials and enables fine-grained sandboxing. It also reduces latency, cutting time-to-first-token by over 90% at P95 in internal measurements.
Event-based interaction model
Instead of simple request-response cycles, the system operates through events, including tool calls, user inputs, and intermediate outputs. This enables real-time streaming, better observability, and detailed logging of agent behavior for debugging and auditing.
Hands-on use case: incident response agent
A demonstration showed how developers can quickly build a site reliability (SRE) agent that analyzes logs, checks deployments, and diagnoses issues such as elevated P99 latency. The agent identified a root cause—database pool exhaustion tied to a specific code change—and suggested remediation steps.
Extensibility with real-world integrations
While the demo used local JSON data, the framework supports integration with production systems like monitoring tools and deployment pipelines. Developers can extend agents to execute fixes, open pull requests, or automate full incident resolution workflows.
Advanced capabilities: memory, sub-agents, and outcomes
Additional features include memory systems for learning from past interactions, sub-agents for parallel task execution, and outcome-based goals that guide agents toward specific results. A “dreaming” mechanism allows agents to refine memory autonomously.
Security and credential management
Built-in vaults securely store credentials with encryption, enabling per-user and per-session access control. This removes the need for custom secret management systems while maintaining strong security boundaries.
Custom infrastructure support
New support for bring-your-own compute allows execution within a developer’s own containers and environments, combining managed orchestration with enterprise infrastructure requirements.
CONCLUSION
Managed agents mark a shift toward fully abstracted AI development, where infrastructure and reliability are handled by the platform, enabling faster deployment of complex, production-grade autonomous systems.