CHAPTERS
Welcome & session roadmap for Claude Managed Agents
Michael and Harrison introduce themselves and set expectations for the talk. They outline a progression: AI capability recap, motivations for Managed Agents, a primitives deep dive, partner discussion of new features, and how to get started.
From Claude 3 to Opus 4.7: capability growth shifts the bottleneck to infrastructure
The speakers describe rapid model improvements—from basic code snippets to agents that can generate merge-ready PRs. They argue the limiting factor is increasingly the surrounding infrastructure (tools, security, long-running tasks) rather than model intelligence.
What stronger agents need: access, identity, and new interaction modes
They explain why agents need secure access to internal systems (credentials, GitHub, Slack/email) plus clear identity/auth. Interaction is moving beyond chat toward outcome-driven tasks and resumable work that can continue across days or months.
Why Anthropic built Claude Managed Agents: memory, infra reliability, observability
Research surfaced recurring pain points that blocked teams from productionizing agents. Context/memory mistakes can derail performance; production demands reliability, scalability, security, and latency; and teams need observability to measure success and diagnose failures.
Core building blocks: define an agent, provision an environment, run a session
They describe the minimal workflow to build with Managed Agents. You define an agent (identity + capabilities), run it inside a sandboxed environment (network allowlists, packages), start a session, and monitor/steer via the event stream.
Event stream explained: user events, agent events, session lifecycle, spans
The event stream is presented as the canonical log of everything that happens in a session. Events are grouped so developers can interpret and build UIs/monitoring around them—covering user inputs, agent actions, session status, and timing boundaries.
Console walkthrough demo: 'Pascal' grocery-habits agent in action
A demo shows an app dashboard integrated with Managed Agents, where an analysis run is launched and observed live. They highlight real-time event visibility, agent configuration display, environment settings, and results surfaced back into the application via API.
Post-run optimization: using session transcripts to improve performance
They show an “Ask Claude” capability where Claude reviews the session transcript and suggests configuration and performance improvements. A concrete example is identifying a long-running uploaded Python script and proposing ways to reduce runtime for better UX.
Getting started paths: Claude Code skill, CLI, and cookbooks
Developers are offered three fast on-ramps to integrate. Claude Code includes an API skill that understands Managed Agents, a CLI enables direct session/agent interaction, and cookbooks provide copy-paste examples across API surfaces.
Advanced features overview: orchestration, outcomes, memory, and dreaming
They highlight higher-level capabilities added recently to move beyond single-thread agents. These include spawning sub-agents for delegation, outcome-based iterative grading loops, long-lived memory stores, and “Dreaming” to distill learning across thousands of sessions.
New launches: self-hosted sandboxes & MCP Tunnels (research preview)
Two new features are introduced to increase power and security in production contexts. Self-hosted sandboxes let teams run tools in their own VPC/compute, while MCP Tunnels securely connect private MCP servers to Claude without exposing them publicly.
Partner panel: different bets on sandbox infrastructure (Vercel, Modal, Daytona, Cloudflare)
Partners explain how their platforms approach sandboxing for agents. Themes include reusable primitives across products (Vercel), massive global scale and GPU support (Modal), human-like dev environments with pause/resume/fork at speed (Daytona), and dual primitives from microVMs to lightweight isolates (Cloudflare).
What partners want built next—and the hard problems ahead (security, identity, resumability)
Partners discuss emerging use cases: personal “chief of staff” assistants, background optimization agents (including GPU-driven profiling/benchmark hill-climbing), and agents operating legacy/closed systems via sandboxes. The conversation turns to ongoing challenges like security boundaries, identity propagation, auth unification, storage/resumability, and collaboration patterns when agents generate most traffic.
