Anthropic’s Managed Agents platform, now in public beta, is a set of composable APIs for running cloud-hosted agents with sandboxed execution, persistent state, long-running sessions, and multi-agent coordination built in. The pitch is that teams currently spend months building the infrastructure around agent calls: credential management, state persistence, checkpointing across disconnections, error recovery, and audit trails. Managed Agents provides those as primitives so teams can focus on the agent behaviour rather than the surrounding plumbing.
The architecture includes an “orchestration harness” that manages tool invocation timing, context handling, and error recovery. Anthropic reports up to 10 percentage point improvements on structured file generation tasks compared to standard prompting loops, which they attribute to the harness enabling tighter iteration and self-evaluation. Agents can spawn and direct other agents (in research preview), and the permissions model uses scoped identities with execution tracing. The effect is that multi-agent pipelines get a degree of governance tooling that was previously custom-built or absent.
This is infrastructure-as-a-service for agent systems, and the competitive dynamic matters. Teams building on this platform are trading control and portability for speed to production. The sandboxed execution environment means you’re not running arbitrary code in your own infrastructure, which has security benefits but also means you’re bound to Anthropic’s execution model. For teams without the engineering bandwidth to build their own agent infrastructure, this collapses months of work. For teams with strong infrastructure opinions or multi-vendor strategies, the lock-in calculus is less favourable.
Early adopters (Notion, Rakuten, Asana, Sentry) reportedly shipped production agents within weeks rather than months. That timeline reflects both the infrastructure savings and the fact that agent product requirements tend to crystallise quickly once the infrastructure friction is gone. The pattern mirrors what cloud compute did to server provisioning: the underlying task is still complex, but the ceiling on how fast you can iterate dropped significantly.
The multi-agent coordination piece is the one to watch. Agents that can spawn and direct subagents is the architectural pattern that makes complex workflows tractable, and it’s also the pattern that makes failure modes harder to reason about. Scoped permissions and execution tracing are the right primitives for that environment. Whether Anthropic’s implementation of those primitives is robust enough for regulated industries or high-stakes workflows is a question early adopters will answer in the next few months.