Building High Performance Agents with Claude Managed Agents on Vercel

Highlights

• Cloud Managed Agents is Anthropic's suite of APIs designed to run AI agents in the cloud.

• Dynamic workflows allow Claude to automatically determine its own multi-agent execution plan and orchestrate up to 1,000 agents.

• Dynamic workflows correctly identified 66 out of 70 embedded code bugs across three test runs, compared to single agents that found between 14 and 27 bugs.

• Vercel sandboxes run on Firecracker micro VMs, providing hardware-level tenant isolation alongside fast startup times.

• Credential brokering injects real authentication tokens at the Vercel sandbox firewall so raw credentials never reside inside the untrusted sandbox code.

Timeline

1. 00:00 - 05:38

   Architecture and Core Problem of Agent Sandboxes

   Building reliable agentic systems requires moving away from spawning inline sandboxes per request, which introduces latency and exposes credentials to untrusted code. By decoupling the durable server harness from short-lived execution environments, system reliability improves significantly while keeping latency down to milliseconds over the standard messages API. Vercel infrastructure supports this model by providing chat SDK user interfaces, Next.js hosting, and webhooks that trigger backend sandbox execution.

   • Cloud Managed Agents provides a dedicated suite of APIs for running cloud-based agents beyond stateless token generation.
   • Booting a fresh cloud code environment inside a sandbox for every incoming request causes reliability risks, latency overhead, and severe security vulnerabilities.
   • Separating the brain from the hands keeps the durable server harness distinct from the untrusted sandbox environment.
   • Vercel infrastructure integrates directly with Cloud Managed Agents to handle frontend interfaces via the chat SDK and execution sandboxes at the backend.
2. 05:38 - 14:16

   Multi-Agent Strategies and Dynamic Workflows

   Single-agent loops eventually degrade during long-running tasks due to laziness, loss of original objectives, and an inability to objectively grade their own outputs. Multi-agent strategies mitigate these failure modes by dividing responsibilities among distinct actors, such as executors, advisors, and grading agents. Dynamic workflows take this a step further by letting Claude write its own multi-phase execution plan, coordinating up to 1,000 individual agents through patterns like fan-out, adversarial verification, and iterative looping.

   • Allocating tokens to specialized jobs such as advising, grading, and dreaming yields higher-quality outcomes than pure single-agent execution.
   • Dynamic workflows enable Claude to autonomously generate deterministic plans and execute multiple agent phases to handle complex projects.
   • Single-agent architectures operating over massive context windows suffer from context rot, self-favoritism, and goal drift over extended runtimes.
   • Dynamic workflow evaluations on a codebase with 70 planted bugs successfully uncovered 66 bugs across three trials.
3. 14:16 - 20:57

   Security Best Practices and Infrastructure Safeguards

   Securing high-volume multi-agent deployments requires strict adherence to infrastructure isolation principles and deny-by-default egress controls. Placing untrusted code inside Firecracker micro VMs prevents tenant breakout issues while maintaining fast execution velocity. Furthermore, credential brokering ensures that raw production API keys and secrets never exist inside the sandbox filesystem where generated code could misuse them.

   • Anthropic enforces security across model alignment, online real-time classifiers, offline automated reviews, and infrastructure responsibility.
   • Isolating monitoring logs and system kill switches from the untrusted sandbox prevents agents from tampering with administrative controls.
   • Vercel sandbox firewalls implement credential brokering to inject placeholder tokens into sandbox code before requests traverse the internet.
   • Firecracker micro VMs provide hardware-level isolation for sandboxes, striking a balance between security and rapid startup performance.