Anthropic has announced a significant expansion to its Claude Managed Agents platform, introducing a suite of features designed to move AI from simple task execution to autonomous self-improvement. The headline feature, “Dreaming,” allows agents to review their own past performance and refine their internal memory between sessions, effectively learning from their own mistakes without human intervention.
This update, detailed on the Anthropic Blog, marks a shift in the competitive landscape. While OpenAI and Google have focused on lightweight orchestration and enterprise data scaling, Anthropic is leaning into the “autonomous execution runtime” model—providing the infrastructure, sandboxing, and now the cognitive feedback loops required for agents to operate over long horizons.
What is Agentic Dreaming?
Despite the poetic name, “Dreaming” is a highly structured, asynchronous process. It acts as a curation layer for an agent’s memory. While standard memory captures incremental data during a live session, Dreaming processes this data offline to maintain a “high-signal” memory store.
According to technical research, the Dreaming workflow follows a four-phase cycle:
- Orientation: The agent establishes its current knowledge state by reading its memory directory.
- Signal Gathering: It scans up to 100 previous session transcripts (JSONL files) to identify recurring patterns, user feedback, and “memory drift.”
- Consolidation: Claude merges new insights, resolves contradictions, and prunes irrelevant noise.
- Generation: The system produces a reorganized memory layer (often a
MEMORY.mdfile) that can be automatically applied or held for developer review.
This isn’t just theoretical. Legal tech firm Harvey reported a 600% increase in task completion rates by using Dreaming to identify tool-specific workarounds and recurring file-type preferences across long-running projects Source.
Outcomes and Multi-Agent Orchestration
Alongside Dreaming, Anthropic released two other features into public beta: Outcomes and Multi-agent Orchestration.
Outcomes introduces a formal evaluation loop. Developers write a rubric describing success, and a separate “grader” agent evaluates the lead agent’s output. Because the grader operates in its own context window, it isn’t influenced by the lead agent’s reasoning chain. In Anthropic’s internal testing, this loop improved task success by up to 10 points, particularly on complex, subjective tasks like matching a brand voice Source.
Multi-agent Orchestration allows a “lead” agent to delegate sub-tasks to specialists. For example, a lead agent investigating a system outage can dispatch sub-agents to parallelize the search through deploy history, error logs, and support tickets. These specialists work on a shared filesystem, and their findings are consolidated back into the lead agent’s context.
The Economics of Managed Agents
Building agents on Anthropic’s infrastructure isn’t just about the model; it’s about the managed runtime. The pricing reflects this “serverless for AI” approach, billing across three dimensions Source:
| Cost Component | Rate |
|---|---|
| Active Runtime | $0.08 per session-hour (billed to the ms) |
| Model Tokens | Standard Claude rates (e.g., $3/$15 per M for Sonnet 3.5) |
| Tool Charges | e.g., $10.00 per 1,000 Web Searches |
Crucially, idle time—such as when an agent is waiting for a human to approve a tool call—is not billed. This makes the platform viable for long-running, asynchronous workflows where the agent might “live” for hours but only execute for minutes.
Competitive Context
Anthropic is positioning Managed Agents as the high-trust, production-grade alternative to the OpenAI Agents SDK. While OpenAI offers high flexibility with “handoff” patterns, Anthropic provides a fully managed sandbox with persistent sessions and native Model Context Protocol (MCP) support.
Compared to Google Vertex AI, which excels at enterprise-scale data integration, Claude Managed Agents are optimized for a “give the agent a computer” paradigm. They are uniquely capable of navigating OS environments and executing bash commands within their managed containers Source.
Implementation & Availability
Managed Agents are currently in public beta. To use the new features, developers must include the managed-agents-2026-04-01 beta header in their API requests.
# Example of initiating a session with the beta header
curl https://api.anthropic.com/v1/managed_agents/sessions \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-beta: managed-agents-2026-04-01" \
-d '{
"agent_id": "agent_abc123",
"environment_id": "env_xyz789"
}'
Takeaways for Practitioners
- Shift to Asynchronous Learning: Dreaming moves agents from “stateless” to “evolving.” Instead of manual prompt engineering to fix recurring errors, you can now let the agent curate its own best practices.
- Quality via Separation: Use the Outcomes feature to separate execution from evaluation. Grader agents are less prone to the “sunk cost” bias of the agent that actually performed the work.
- Infrastructure over Orchestration: Managed Agents solve the “boring” parts of AI—sandboxing, credential vaulting, and session persistence—allowing teams to move from prototype to production in days Source.
- Watch the Tool Costs: While token costs are falling, tool-specific charges (like $10/1k searches) can quickly become the dominant line item in an agentic budget.