How to Supercharge Your AI Agents with Anthropic's Managed Agent Platform

Introduction

Anthropic's Managed Agents platform now offers three powerful features that transform your AI agents from simple task executors into self-improving, quality-driven collaborators. With dreaming, agents review their past work to find patterns and correct mistakes automatically. Outcomes let you define what 'good' looks like, with a separate grader agent evaluating results. And multi-agent orchestration enables breaking down complex tasks into parallel sub-tasks handled by multiple agents. This step-by-step guide shows you how to configure and use these features in your own projects, making your agents more autonomous and effective with minimal manual steering.

What You Need

• An active Anthropic account with access to Managed Agents (available in public beta or later versions)
• Basic familiarity with the Anthropic console or API for configuring agents
• A use case where agents perform recurring tasks (e.g., customer support, data processing, content generation)
• Optional: Multiple tasks that can be parallelized for orchestration

Step-by-Step Guide

Step 1: Set Up Your Managed Agent Environment

Before diving into the advanced features, ensure your base agent is properly configured. Log in to the Anthropic console and create a new Managed Agent. Define its primary goal, such as handling customer inquiries or processing support tickets. Provide initial instructions and a knowledge base if needed. This foundation will allow you to layer on dreaming, outcomes, and orchestration.

Step 2: Enable Memory and Configure Dreaming

The dreaming feature builds on persistent memory. In your agent settings, activate memory to store information across sessions. Then, under the advanced options, enable dreaming (currently in research preview). You can choose between automated dreaming (where the agent runs scheduled reviews without human intervention) or manual review (where you approve memory updates). Dreaming analyzes recent sessions, identifies patterns—including mistakes—and updates the memory with holistic observations. This self-improvement loop helps the agent learn from its own work, similar to how humans consolidate memories during sleep.

Step 3: Define Outcomes for Quality Control

Outcomes allow you to set explicit success criteria for your agent. In the agent configuration, navigate to the Outcomes section. Create a set of criteria that define what 'good' looks like for your task—for example, response accuracy, tone compliance, or completeness. Anthropic automatically sets up a separate grader agent that evaluates each output against these criteria using its own context window. This prevents the main agent from 'cheating' and ensures objective assessment. Use outcomes for tasks requiring attention to detail or subjective quality, like maintaining brand voice in marketing copy. Anthropic reports that using outcomes improves task success by up to 10 points over standard prompting.

Step 4: Orchestrate Multiple Agents for Complex Tasks

When facing a large, multi-step project, enable multi-agent orchestration. In your project dashboard, create a parent agent that can delegate sub-tasks. For instance, a customer support system might involve one agent for ticket triage, another for technical answers, and a third for escalation handling. Configure the parent to break down the main objective into sub-tasks, then assign each to a specialized agent. You can control parallel execution and consolidation of results. This feature is ideal for workflows that benefit from parallel processing, such as data analysis, content generation at scale, or complex decision trees.

Conclusion Tips

• Start small: Test dreaming with a single agent on a low-impact task before rolling out to production.
• Review initial dreams: Even in automated mode, occasionally inspect memory updates to ensure the agent's patterns are correct.
• Iterate on outcomes: Refine your criteria based on grader feedback to continuously improve quality.
• Balance orchestration: Too many agents can create overhead; use orchestration only when tasks are clearly parallelizable.
• Monitor performance: Track success rates and error patterns to fine-tune memory, dreaming frequency, and outcome thresholds.

By following these steps, you can turn your Managed Agents into a self-improving, quality-focused, and scalable workforce. The combination of dreaming for pattern recognition, outcomes for quality assurance, and orchestration for parallel work gives you a robust system that handles complex tasks with minimal human steering.