An Anthropic engineer demonstrated Claude Managed Agents, a server-side harness for production agents. The demo included an SRE agent that traced a P99 latency spike to a specific commit.
Key facts
- 90%+ lower P95 time-to-first-token claimed
- SRE agent traced P99 latency spike to exact commit
- Sessions survive refreshes for long-running workflows
- Includes encrypted credential vaults and webhook triggers
- Bring-your-own-compute option for enterprise data control
Claude Managed Agents, unveiled in a demo by an Anthropic engineer, is a new harness for running AI agents server-side without managing hosting or scaling. According to @_vmlops, the architecture separates agents (the brain), environments (the hands), and sessions (the glue), with the agent loop executing server-side while tool calls are decoupled. Sessions persist across refreshes, a key requirement for long-running production workflows.
The demo highlighted several production-ready features: a 90%+ reduction in P95 time-to-first-token, bring-your-own-compute options, encrypted credential vaults, sub-agents, memory, outcomes tracking, and webhook triggers. The company did not disclose specific latency numbers or pricing for the managed service.
What the live demo showed
The centerpiece was an SRE agent built on the harness that investigated a P99 latency spike in real time. It analyzed metrics, deployments, and diffs, then traced the incident to the exact commit that caused it. This is a meaningful step beyond typical agent demos, which often focus on code generation or retrieval rather than operational debugging.
Why this matters for production agents
The shift to server-side execution addresses a common pain point: agents that lose state or require heavy client infrastructure. By decoupling tool execution and persisting sessions, Claude Managed Agents positions itself as a platform for continuous, autonomous operation. The bring-your-own-compute option is notable, suggesting Anthropic is targeting enterprises that need to keep data on their own infrastructure.
Anthropic has not yet announced general availability or pricing for Claude Managed Agents. The demo, however, signals a push to own the agent runtime layer, competing with similar offerings from OpenAI and open-source frameworks like LangGraph.
What to watch
Watch for Anthropic's official launch announcement of Claude Managed Agents, including pricing and GA timeline. Key metrics to track: adoption of bring-your-own-compute among enterprises, and whether the 90% P95 latency improvement holds in third-party benchmarks. Also monitor for competitive responses from OpenAI's agent platform.
