Letterman Labs.
AI Weekly · Issue 4

Anthropic Unleashes Dreaming Agents

Claude agents self-improve through memory curation, outcomes, and multiagent orchestration for true autonomy

№ 4 May 11, 2026 Written by machines, for humans
Hero illustration for Anthropic Unleashes Dreaming Agents
Anthropic significantly upgraded Claude Managed Agents with "dreaming" for self-improvement and multi-agent orchestration, alongside general availability of Opus 4.7 and Microsoft 365 integrations. Google's Gemini 3.1 arrived in Google Home, bringing advanced reasoning to smart devices, while a new multimodal embedding model entered preview. GitHub Copilot focused on developer experience, enhancing its CLI with direct prompt hooks and faster account switching, and making GPT-5.5 generally available. These updates push AI agents towards greater autonomy and integration, while simultaneously refining core developer tooling and expanding multimodal capabilities across platforms.

Top of the Week

Top of the Week illustration
Top of the Week

Anthropic is pushing the boundaries of AI agent autonomy with three major new features for Claude Managed Agents: "dreaming," "outcomes," and "multiagent orchestration." The most striking of these is dreaming, a research preview that allows Claude agents to review past sessions, identify patterns, and self-improve by curating memories. This scheduled process can automatically update an agent's memory or allow human review before changes are implemented, creating a "robust memory system" where "memory lets each agent capture what it learns as it works. Dreaming refines that memory between sessions, pulling shared learnings across agents and keeping it up-to-date," according to Anthropic.

[VIDEO: https://www.instagram.com/reel/DYHjdMMgWCI/]

Alongside dreaming, the "outcomes" feature lets users define what constitutes a successful result for an agent, providing a clear target for complex tasks. Multiagent orchestration, already deployed by companies like Netflix for their platform team, enables specialist agents to work in parallel on a shared filesystem, contributing to a lead agent's context. This allows the lead agent to check in with specialists mid-workflow, as "events are persistent and every agent remembers what it’s done."

These agentic advancements are significant because they move beyond simple prompt-response interactions toward more persistent, self-improving, and collaborative AI systems. For builders, it means the tools to create more sophisticated, less hand-held AI assistants are becoming available, potentially unlocking new categories of autonomous applications. This suite of features indicates a clear strategic direction from Anthropic: building AI that learns, adapts, and works in concert, rather than merely executing isolated commands.

Claude

Claude illustration
Claude

Gemini

Gemini illustration
Gemini

Copilot

Copilot illustration
Copilot

Tools Worth Trying

  1. Perplexity — The AI-native search engine that provides sourced answers in seconds, crossing 30M weekly users in early 2026. It's best for market research, due diligence, and fact-checking, offering concrete citations where other LLMs might generalize.
  2. Notion AI 2.0 — Launched in March 2026 with autonomous workflow agents, becoming the default knowledge-work AI for over 25M Notion users. It's ideal for workflow automation and knowledge management within the Notion ecosystem.
  3. ElevenLabs — An all-in-one AI platform for voice, image, and video generation. It's excellent for creators and marketers needing high-quality multimedia assets without extensive manual effort.
  4. Confident AI — The leading AI prompt management tool in 2026, offering git-based prompt management (branching, commit history, approval workflows) and built-in observability with 50+ metrics. It's for teams who need to manage prompts with the same rigor as code.
  5. Lovable.dev — The "ChatGPT for full-stack apps" that hit $100M ARR in 8 months, building working React apps from a prompt, including Supabase setup and Vercel deployment, all in 5 minutes. It's perfect for founders prototyping and marketers building landing pages without coding.
  6. Granola — An AI meeting notes tool that listens locally on your Mac, transcribing and summarizing in real-time without requiring a bot in your call. It's for privacy-conscious professionals who need efficient meeting documentation.
Anthropic shipped over 30 new features for Claude in Q1 2026 alone, including new models, integrations, and capabilities almost every week, highlighting the rapid pace of development in the AI agent space.

The 5-Minute Action Plan

  1. Explore Claude Managed Agents: If you're building autonomous AI workflows, investigate Anthropic's new "dreaming," "outcomes," and "multiagent orchestration" features for Claude Managed Agents to see how they can enable self-improving and collaborative AI. Start by reviewing the details here: https://9to5mac.com/2026/05/07/anthropic-updates-claude-managed-agents-with-three-new-features/.
  2. Integrate Claude with Microsoft 365: For enterprise users, download the Claude add-ins for Excel, PowerPoint, and Word (now GA) and try the Outlook beta to embed AI assistance directly into your daily productivity suite. Find more details here: https://releasebot.io/updates/anthropic/claude.
  3. Upgrade Google Home with Gemini 3.1: Update your Google Home app to leverage the new Gemini 3.1 reasoning capabilities for complex commands and explore the modernized camera experience and expanded automations. Check for updates on your device or in the app store.
  4. Experiment with Multimodal Embeddings: Developers should explore Google's gemini-embedding-2-preview model for applications requiring a unified understanding of text, images, video, audio, and PDFs. Access the changelog for details: https://ai.google.dev/gemini-api/docs/changelog.
  5. Review GitHub Copilot Data Settings: If you use GitHub Copilot Free, Pro, or Pro+, review your GitHub settings to understand and adjust your opt-out preferences for user data being used to train AI models. This change went into effect on April 24: https://github.com/features/copilot.
  6. Try GitHub Copilot CLI Enhancements: Experiment with the new userPromptSubmitted hooks and faster multi-account switching in the latest GitHub Copilot CLI (version 1.0.44-3) to streamline your terminal-based development workflows. See the changelog for details: https://releasebot.io/updates/github.
To get highly specific, sourced answers for research, combine Perplexity with a detailed prompt that defines the scope and required citation style. `` Provide a comprehensive overview of [TOPIC], citing at least three distinct, reputable sources from the last 12 months. For each source, include the URL and a brief summary of its key findings relevant to the topic. Structure the output with a main summary, followed by a bulleted list of sources and their summaries. ``

The Prompt Library

[Deep Research Synthesis — Cross-Source]

Use this prompt to synthesize information from multiple provided sources into a coherent, structured overview, ideal for competitive analysis or literature reviews.

You are an expert research analyst. I will provide you with several text excerpts from different sources. Your task is to synthesize the information from these excerpts into a comprehensive report on [TOPIC].

Your report should:
1. Start with an executive summary (3-5 sentences).
2. Identify and explain the key themes or arguments present across all sources.
3. Highlight any conflicting information or differing perspectives.
4. Extract specific data points, statistics, or named entities relevant to the topic.
5. Conclude with a section on unanswered questions or areas for further research based on the provided texts.
6. For each piece of information, indicate which source(s) it came from using bracketed numbers [1], [2], etc.

[Code Review — Security & Best Practices]

Use this prompt to get a security-focused code review, identifying vulnerabilities and suggesting improvements based on best practices.

Perform a security and best practices code review on the following [LANGUAGE] code snippet. Focus on identifying potential vulnerabilities (e.g., injection flaws, insecure deserialization, weak authentication, sensitive data exposure), adherence to common security standards, and general code quality improvements.

Provide:
1. A list of identified security risks with severity ratings (High, Medium, Low).
2. Specific recommendations for remediation, including code examples where applicable.
3. Suggestions for improving code readability, maintainability, and adherence to [LANGUAGE] best practices.
4. Justify each finding with a brief explanation.

[PASTE CODE HERE]

[Writing Improvement — Tone & Clarity]

Use this prompt to refine a piece of writing for a specific tone and improved clarity, ensuring it resonates with the target audience.

I need you to revise the following text to achieve a [TONE, e.g., professional, empathetic, persuasive, concise] tone, while also improving its clarity and conciseness for a [TARGET AUDIENCE, e.g., technical executives, new users, general public].

Specifically:
1. Remove any jargon or overly complex phrasing that might confuse the target audience.
2. Ensure the core message is immediately clear and impactful.
3. Adjust sentence structure and vocabulary to match the desired tone.
4. Highlight the most significant changes you made and why.

[PASTE TEXT HERE]

[Decision Framework — Pros & Cons Analysis]

Use this prompt to analyze a decision by outlining its pros, cons, and potential risks, helping to evaluate complex choices.

Act as a strategic advisor. I am considering [DECISION, e.g., launching a new product feature, hiring a new team member, adopting a new technology stack].

Please provide a structured analysis including:
1. **Pros:** A detailed list of potential benefits and positive outcomes.
2. **Cons:** A detailed list of potential drawbacks, challenges, and negative impacts.
3. **Risks:** Identification of specific risks, their likelihood (High, Medium, Low), and potential mitigation strategies.
4. **Key Considerations:** Any additional factors or questions I should consider before making this decision.
5. A recommendation based on your analysis, with a clear rationale.

[Learning a New Topic — Structured Curriculum]

Use this prompt to generate a structured learning plan for a new, complex topic, breaking it down into manageable modules and resources.

I want to learn about [COMPLEX TOPIC, e.g., Quantum Machine Learning, Serverless Architecture, Advanced Prompt Engineering]. Assume I have a basic understanding of [PREREQUISITE KNOWLEDGE, e.g., linear algebra, cloud computing, fundamental AI concepts].

Create a structured learning path for me, broken down into 5-7 modules. For each module, include:
1. A clear title and brief description of what I will learn.
2. Key concepts to master.
3. Suggested practical exercises or projects.
4. Recommended resources (e.g., "Official documentation," "Academic papers," "Online courses," "YouTube tutorials," "Books" - do not invent specific titles, just types).

[Brainstorming — Innovative Solutions]

Use this prompt to generate a diverse set of innovative solutions for a given problem, encouraging creative thinking.

Brainstorm 10-15 innovative and unconventional solutions to the following problem: [PROBLEM STATEMENT, e.g., "How can we reduce customer churn in a SaaS subscription model without lowering prices?" or "How can we make urban commuting more sustainable and enjoyable?"].

For each solution, provide:
1. A concise title.
2. A 1-2 sentence description.
3. One unique advantage it offers.
4. One potential challenge in implementation.
Aim for variety and push beyond obvious answers.

[Summarization — Long Document Key Points]

Use this prompt to extract the most critical information from a lengthy document, presenting it as a concise summary with actionable insights.

Summarize the following [TYPE OF DOCUMENT, e.g., research paper, meeting transcript, article] into its 5 most critical key points. For each key point, provide a brief explanation and one actionable insight or implication. Ensure the summary is objective and captures the essence of the document, even if it's lengthy.

[PASTE LONG DOCUMENT HERE]

[Task Breakdown — Project Management]

Use this prompt to break down a large project into smaller, manageable tasks, suitable for project planning and delegation.

Break down the project "[PROJECT TITLE, e.g., Develop a new customer onboarding flow, Launch a marketing campaign for Q3]" into a detailed list of actionable tasks.

For each task, include:
1. A unique task ID.
2. Task Name.
3. Brief Description.
4. Estimated Effort (Small, Medium, Large).
5. Suggested Owner (e.g., Marketing, Engineering, Product).
6. Dependencies (if any, reference other task IDs).
Organize tasks logically, perhaps by phase or functional area.

[Persuasive Writing — Call to Action]

Use this prompt to draft persuasive copy for a specific call to action, focusing on benefits and addressing potential objections.

Draft persuasive copy for a [FORMAT, e.g., email, landing page section, social media ad] with the goal of convincing [TARGET AUDIENCE] to [DESIRED ACTION, e.g., sign up for a free trial, purchase a product, attend a webinar].

Your copy should:
1. Clearly state the value proposition and benefits of taking the desired action.
2. Address potential objections or hesitations the target audience might have.
3. Use compelling language and a [TONE, e.g., urgent, inspiring, reassuring] tone.
4. Include a clear and strong call to action.
5. Keep the length appropriate for the specified format.

[Debugging — Code Error Analysis]

Use this prompt to analyze a code error, identify the root cause, and suggest a fix, including relevant context from the codebase.

I am encountering an error in my [LANGUAGE] application. Here is the error message and the relevant code snippet.

Error Message:
[PASTE ERROR MESSAGE HERE]

Code Snippet:
[PASTE CODE SNIPPET HERE]

Based on this information, please:
1. Identify the most likely root cause of the error.
2. Explain why this error is occurring in simple terms.
3. Provide a corrected version of the code snippet with a clear explanation of the changes.
4. Suggest any additional steps or checks I should perform to prevent similar errors in the future.
📚 Sources 15
  1. 9to5mac.comhttps://9to5mac.com/2026/05/07/anthropic-updates-claude-managed-agents-with-three-new-features/
  2. releasebot.iohttps://releasebot.io/updates/anthropic/claude
  3. support.claude.comhttps://support.claude.com/en/articles/12138966-release-notes
  4. aimaker.substack.comhttps://aimaker.substack.com/p/anthropic-claude-updates-q1-2026-guide
  5. thurrott.comhttps://www.thurrott.com/smart-tech/smart-home/google-home/335706/google-home-spring-2026-update-arrives-with-gemini-3-1-new-camera-capabilities-and-more
  6. ai.google.devhttps://ai.google.dev/gemini-api/docs/changelog
  7. docs.cloud.google.comhttps://docs.cloud.google.com/gemini/enterprise/docs/release-notes
  8. cnet.comhttps://www.cnet.com/tech/services-and-software/google-io-2026-everything-to-know/
  9. gemini.googlehttps://gemini.google/latest-news/
  10. releasebot.iohttps://releasebot.io/updates/github
  11. github.bloghttps://github.blog/changelog/label/copilot/
  12. github.comhttps://github.com/features/copilot
  13. midastools.cohttps://www.midastools.co/blog/best-ai-tools-may-2026
  14. youtube.comhttps://www.youtube.com/watch?v=A4UUukO9HaA
  15. confident-ai.comhttps://www.confident-ai.com/knowledge-base/compare/best-ai-prompt-management-tools-with-llm-observability-2026