Code with Claude 2026: Managed Agents, Dreaming & AWS

Anthropic held its first developer conference, Code with Claude, at SVN West in San Francisco on May 6, 2026. The event covered a lot of ground — but the throughline was clear: Anthropic is building a full platform layer for agents, not just a model API.

This article breaks down every announcement in terms developers can act on today.

The Conference at a Glance

Code with Claude 2026 ran as a three-city tour: San Francisco on May 6, London on May 19, and Tokyo on June 10. The SF keynote covered four main tracks: the opening keynote, a "What's new in Claude Code" session, a "Building on Claude at GitHub scale" talk, and a "Get to production faster with Managed Agents" session.

The biggest theme was the shift from prompt engineering to agent engineering. Anthropic made the case that developers who only use Claude as a one-shot code generator are leaving most of the value on the table. The platform they announced is designed to support long-running, self-improving agents that coordinate in parallel and get smarter between jobs.

Claude Managed Agents: The Platform

Before the conference announcements make sense, it helps to understand the base layer. Claude Managed Agents launched in public beta on April 8, 2026 — about a month before Code with Claude. It is Anthropic's hosted runtime for stateful, persistent agents.

Key facts:

• Beta header: managed-agents-2026-04-01 (set automatically if you use the SDK)
• Supported models: claude-opus-4-7 and claude-sonnet-4-6
• Pricing: standard API token rates for whichever model the agent uses, plus a session runtime charge of $0.08 per session-hour
• Features at launch: memory, tools, MCP connectors, Skills, web search, code execution, batch processing

The three features announced at Code with Claude — Dreaming, Outcomes, and multiagent orchestration — are all addons to this base platform.

Dreaming: Agents That Improve Between Sessions

Dreaming is the most unusual feature Anthropic has shipped in some time. It is a scheduled background process that runs between agent sessions, reviews what the agent did, extracts patterns from the session logs and memory stores, and writes new memory entries the next session can use.

The name is intentional. Anthropic explicitly draws the analogy to hippocampal memory consolidation — the process by which the human brain replays the day's events during sleep and decides what to keep. The parallel is closer than it sounds: dreaming is not inference, it is a compression and curation job that runs asynchronously, after the session is over.

Why It Matters

The failure mode dreaming is designed to fix is subtle but expensive: agents forget things between sessions. An agent that does legal document review might encounter a quirk with a particular filetype on Monday, work around it, and then hit the exact same failure on Tuesday because the workaround was never written to memory. Over time this adds up.

Legal-AI startup Harvey ran dreaming in internal testing before the public launch. According to Anthropic's official announcement, Harvey's task completion rates rose roughly 6x. The use case was specific: Harvey's agents were repeatedly hitting the same file-handling failures and forgetting the fix. Dreaming consolidated those workarounds into persistent memory, and the failure rate dropped.

Harvey's numbers are self-reported through Anthropic's announcement, not an external benchmark. Teams building on different workloads should run their own baselines.

Developer Access

Dreaming is in research preview — it requires a separate access request and an additional beta header:

anthropic-beta: dreaming-2026-04-21

To configure a dreaming schedule in your agent definition:

name: Legal Document Reviewer
model: claude-opus-4-7
system: You review legal documents for compliance issues.
dreaming:
  enabled: true
  schedule: "0 2 * * *"   # runs nightly at 2am UTC
  max_sessions_to_review: 50

The schedule field accepts standard cron syntax. Anthropic recommends off-peak hours since dreaming adds a small background cost at standard token rates.

Outcomes: Rubric-Driven Iteration

Outcomes is in public beta and is a simpler feature to explain: you write a rubric that defines success, and Claude works until it meets it.

The mechanism has two parts. First, the agent produces output. Then, a separate grader agent evaluates that output against your rubric — in its own context window, completely isolated from the agent's reasoning. If the grader identifies a problem, it tells the agent specifically what went wrong, and the agent takes another pass.

Why Isolated Grading Matters

Having the grader run in a separate context prevents a common failure mode in self-critique loops: the model sees its own reasoning and rationalizes instead of genuinely evaluating. By keeping the grader context clean, Outcomes gets a more honest assessment.

Rubric Example

A rubric is plain text. You pass it in the agent configuration:

name: Code Review Agent
model: claude-sonnet-4-6
outcomes:
  rubric: |
    The review must:
    - Identify all security vulnerabilities in the diff
    - Suggest specific fixes with code examples
    - Not flag style issues as blocking
    - Produce output under 500 words
  max_iterations: 3

The max_iterations field caps retries so you do not burn unbounded tokens on a task that is not converging.

Multiagent Orchestration: Parallel Specialists

The third announcement is multiagent orchestration, also in public beta. A coordinator agent can decompose a task and delegate subtasks to specialist agents — each with its own model, system prompt, tools, and independent context window.

The practical example from the conference: an incident-response agent that receives an alert, then fans out in parallel to four specialists: one combing deploy history, one scanning error logs, one reading metrics, one reviewing support tickets. All four run simultaneously on a shared filesystem and report back to the coordinator.

YAML Configuration

name: Engineering Lead
model: claude-opus-4-7
system: >
  You coordinate engineering work. Delegate code review
  to the reviewer agent and test writing to the test agent.
tools:
  - type: agent_toolset_20260401
multiagent:
  type: coordinator
  agents:
    - type: agent
      id: $REVIEWER_AGENT_ID
    - type: agent
      id: $TEST_WRITER_AGENT_ID

Key limits: the coordinator can list up to 20 unique agents in multiagent.agents. It can call multiple copies of each, and it can delegate to a copy of itself using type: self for recursive decomposition.

At the conference, Anthropic also announced rate limit increases tied to multiagent workloads. Pro and Max subscription plan limits were doubled as of the keynote date.

Agent View in Claude Code

Alongside the Managed Agents announcements, Anthropic shipped Agent View in Claude Code as a research preview on May 11, 2026. This is a terminal-native dashboard that lists every running Claude Code session in one pane.

It ships with Claude Code v2.1.139 or later and requires a Pro, Max, Team, Enterprise, or Claude API plan.

How to Use It

Three ways to open agent view:

1. Run claude agents from any terminal
2. Press ← on an empty prompt inside any existing session
3. Run /bg inside a session to background it and return to agent view automatically

Once inside agent view, you can select any row with the arrow keys and press Space to open a peek panel. The peek panel shows either the session's most recent output or the question it is waiting on — without opening the full transcript. You can reply directly from the peek panel.

To launch a headless background session that never opens a foreground terminal:

claude --bg "Refactor the payments module to use the new Stripe API"

Background sessions still appear in agent view and still surface when they need input.

Why This Is Useful

Before Agent View, running multiple Claude Code sessions meant juggling multiple terminal tabs with no shared status. If a background session hit an error or needed a decision, you would not know until you switched to that tab. Agent View makes parallel work manageable without switching contexts constantly.

The conference session on Claude Code also covered CI integration: claude review for PR code review (used internally at Anthropic across all teams) and CI auto-fix, which opens automated fixes against pull requests. These are separate from Agent View but part of the same "Claude as teammate" direction.

Claude Platform on AWS: What Actually Changed

The final major announcement was the general availability of Claude Platform on AWS, with AWS as the first cloud provider to offer the native Claude Platform experience.

There are now three ways to run Claude on AWS:

1. Claude on Amazon Bedrock — AWS-operated, data stays inside the AWS boundary, available via AWS SDK
2. Claude Platform on AWS — Anthropic-operated, billed through AWS Marketplace, IAM authentication
3. Anthropic direct API — separate billing, no AWS integration

Per-token rates are identical across all three. The differences are in billing mechanics, feature availability, and data residency.

What Claude Platform on AWS Adds

The critical functional difference: as of May 2026, Amazon Bedrock does not ship Claude Managed Agents, the full Skills system, native code execution, native web search, or the Anthropic Console with evaluations. Claude Platform on AWS ships all of these on day one and gets new features at the same time as the direct API.

For teams on an AWS Enterprise Discount Program (EDP), Claude Platform on AWS lets EDP commitments retire against Anthropic usage. That matters for organizations that have pre-committed AWS spend.

Authentication Mechanics

import anthropic
import boto3

# Claude Platform on AWS uses SigV4 — handled by the SDK
client = anthropic.AnthropicAWS(
    aws_region="us-east-1",
    # IAM credentials from environment or instance profile
)

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}]
)

CloudTrail audit logging is included at no extra cost. IAM-based access control means you can use existing AWS IAM policies to scope which teams or services can call which Claude features.

Putting It Together: What to Do This Week

The announcements at Code with Claude are not a future roadmap — most of these features are available right now.

If you already use Claude Managed Agents: enable Outcomes and set a rubric for your most error-prone tasks. It is public beta, zero additional configuration required beyond the rubric YAML.

If you run multi-step workflows that could parallelize: try the multiagent orchestration config. A coordinator agent with two or three specialists on parallel data-gathering tasks can cut wall-clock time significantly on tasks that currently run sequentially.

If your agents repeat the same failures: apply for the Dreaming research preview access. The signup form is linked from the official announcement. Harvey's numbers are specific to legal drafting, but the underlying pattern — agents forget workarounds between sessions — is common across long-running agentic workflows.

If your organization is on AWS: evaluate Claude Platform on AWS versus your existing Bedrock setup. The deciding question is whether you need Managed Agents and the full Skills system, or whether data-residency requirements mandate Bedrock. For most teams building new agentic features, Claude Platform on AWS is the more capable starting point today.

For Claude Code users: upgrade to v2.1.139 and try Agent View. Even running two parallel sessions and watching them from a single pane is qualitatively different from tab-switching.

FAQ

Q: Is Dreaming available without a separate beta access request?

Not yet. As of May 2026, Dreaming requires both a manual access request form submission and the dreaming-2026-04-21 beta header. The base Managed Agents beta (managed-agents-2026-04-01) does not unlock Dreaming automatically.

Q: Does multiagent orchestration cost more than single-agent sessions?

There is no orchestration surcharge. You pay standard token rates for each agent plus $0.08 per session-hour per running session. Running four parallel specialists will cost four times the session-hour charge while they are active, but the wall-clock time is typically much shorter.

Q: Can Bedrock customers use multiagent orchestration?

As of May 2026, Amazon Bedrock does not support Claude Managed Agents. Multiagent orchestration is a Managed Agents feature. Teams on Bedrock need to switch to Claude Platform on AWS or the direct Anthropic API.

Q: What models work with Managed Agents?

Claude Managed Agents supports claude-opus-4-7 and claude-sonnet-4-6. The coordinator and each specialist can use different models — a common pattern is using claude-opus-4-7 for the coordinator and claude-sonnet-4-6 for high-throughput specialists.

Q: Where does Agent View data live?

Agent View is a local terminal UI — it reads session state from the Claude Code process running on your machine. Session data for Managed Agents is handled by Anthropic's platform or (for Claude Platform on AWS) routed through Anthropic's infrastructure with IAM logging.

Key Takeaways

Code with Claude 2026 marked a shift in how Anthropic is positioning Claude for developers: not as a model you call once, but as a platform you build persistent agents on. The Dreaming, Outcomes, and multiagent orchestration features are each solving a concrete class of production agent failure. Claude Platform on AWS removes the feature gap between the direct API and AWS-native deployments. Agent View makes parallel Claude Code sessions practical.

The features that require a waiting period (Dreaming research preview access) are worth requesting now. The features in public beta (Outcomes, multiagent orchestration) are production-ready.