Claude

How to get to production faster with Claude Managed Agents

Building agents used to mean spending development cycles on secure infrastructure, state management, permissioning, and reworking your agent loops for every model upgrade. Managed Agents, on the Claude Platform, now handles that layer for you. This session covers how to build and deploy a production-grade agent at scale.

May 6, 2026·17m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Why agent development is shifting from prompts to runtimes

   Jess frames how rapid model improvements are expanding what we ask agents to do—from short, tightly guided tasks to overnight and eventually multi-week projects. As task horizons grow, the limiting factor becomes infrastructure, not model intelligence, which motivates a more robust agent runtime.

2. 2

   Long-horizon requirements: reliability, security, and outcome-oriented work

   Lance explains why long-running agents raise the bar on reliability and security, since failures or leaks become more costly over days or weeks. He introduces outcome-oriented tasks and the need for agents to pause/resume and ask clarifying questions during execution.

3. 3

   Developer pain points that blocked production adoption

   Jess shares pre-launch research showing where developers struggled most building agents themselves. The biggest issues were context management, infrastructure/credentials/security concerns, and lack of observability for probabilistic systems.

4. 4

   What Claude Managed Agents provides: infra + building blocks + observability

   Jess positions Managed Agents as an end-to-end platform that bundles the runtime concerns with composable primitives. The goal is to make agents easier to build, safer to run, and measurable through a “single pane of glass” observability experience.

5. 5

   Core mental model: Agents, Environments, Sessions, and Events

   Lance explains the conceptual building blocks: an agent is a configuration (model, prompt, tools, skills), an environment is the sandbox it operates in, and a session is a single execution with resources and goals. Sessions emit events that you can use to monitor and integrate agent behavior.

6. 6

   Event topology for complex agents: user, agent, session, and span events

   Jess breaks down how Managed Agents categorizes emitted events so developers can reason about complex long-horizon executions. This event model supports steering, lifecycle tracking, and grouping related activity for analysis.

7. 7

   Demo 1 — “Pascal” analytics agent with real-time console tracing

   Jess demos a sample agent, Pascal, running against a hypothetical grocery dataset to generate analytics quickly using a preloaded Python environment. The console shows real-time event updates plus configuration and environment details, enabling after-the-fact diagnosis.

8. 8

   Debugging and improving runs using console analytics

   After the Pascal run completes, the full event stream is available for inspection. The console includes debugging assistance to find bottlenecks, recommend actions, and feed improvements back into development workflows (e.g., Claude Code).

9. 9

   Getting started tooling: Claude Code skill, CLI workflows, and cookbooks

   Lance outlines practical entry points: a Claude Code skill (via /claude-api), a powerful CLI for YAML-based agent configs and session retrieval, and cookbooks for patterns. He emphasizes using tooling to generate code and analyze logs rather than writing everything manually.

10. 10

    Advanced capabilities: orchestration, outcomes, memory, and “dreaming”

    Jess highlights newer features extending agent experiences: multi-agent delegation, outcomes (rubrics/exit criteria), persistent memory across sessions, and a “dreaming” platform where agents reflect and codify learnings. Together these capabilities enable higher fidelity, iterative improvement, and long-term adaptation.

11. 11

    Demo 2 — “Boss Agent” dashboard: outcomes + multi-agent for fast visualizations

    Lance demonstrates a CEO-style interface that answers questions using fake org data and renders artifact-like visualizations in a browser. Outcomes act as a rubric-driven evaluator loop (including screenshot/timing checks), and multi-agent parallelism accelerates multi-chart rendering.

12. 12

    Performance optimization loop: inner outcomes + outer human feedback

    Lance explains a two-loop iteration model: outcomes provide an automated inner loop for rubric-based refinement, while an outer loop lets a user review results and use Claude Code to adjust prompts/rubrics and rerun sessions. He shows measured speedups achieved autonomously (e.g., parallel tool calls, fast mode, prompt optimization).

13. 13

    Closing: ecosystem partners, resources, and how to start building

    Jess closes by crediting user feedback and highlighting partners like Asana and Notion building on the platform. The talk ends with pointers to developer docs and an interactive quickstart to build an agent quickly, plus an invitation to keep sharing feedback.