Top story
Anthropic shipped three Claude Managed Agents features at Code with Claude on May 6-7: Dreaming (offline memory consolidation between runs), Outcomes (rubric-based self-grading with auto-retry), and Multiagent Orchestration (parallel specialist sub-agents under one supervisor). Named numbers from stage: Harvey legal agents at 6x task completion, Wisedocs at 50% document-review-time reduction, Netflix scanning build logs from hundreds of repos in parallel. Dreaming is in research preview; Outcomes and Orchestration are public beta.
Why it matters. If you build on Cursor, Cline, or Aider, the model powering those tools just got a self-improvement loop you can trigger from the API. Read our full writeup.
Quick hits
Mastra hit 23.9k stars and shipped v1.33.0.
The TypeScript agent framework now routes across 40+ model providers, adds human-in-the-loop checkpoints, and has built-in memory management. If your team already writes TS and wants off LangGraph, this is the clearest path. Repo.
DeepSeek-TUI crossed 20.8k stars in a week.
A Rust-native coding agent for DeepSeek models that runs entirely in your terminal. No browser, no Electron wrapper. Useful if you want local-model agent runs without the GUI overhead. Repo.
Perplexity Comet reached Enterprise availability.
Model picker is live for Max subscribers: Opus 4.6 by default, Sonnet 4.5 as an option. The agentic-browser category now has a credible second player beyond pure search. Comet page.
AIDC-AI/Pixelle-Video gained 4.3k GitHub stars this week.
Automated short-video generation engine: script, visuals, and voiceover from a single prompt. Open-source Python, self-hostable. Repo.
Tools to try this week
Cloudways.
Managed cloud hosting built for spinning up self-hosted AI tools: Open WebUI, n8n, or your own model server. Five minutes from signup to a running instance. Try Cloudways.
CustomGPT.
No-code GPT trained on your own docs. Good fit for customer-support chat over a product knowledge base without writing retrieval logic from scratch. Try CustomGPT.
ElevenLabs.
Voice AI for narration, podcast intros, and agent TTS. Low-latency streaming API if you're wiring audio output into an agent pipeline. Try ElevenLabs.
From the Pondero stack
Our local LLM stack guide went up today: Ollama vs LM Studio vs Open WebUI on an M4 Mac mini with Llama 3.3 70B Q4. The short answer is that they are not competitors. Ollama runs the model, LM Studio handles prompt experiments, Open WebUI gives your team the chat surface. Read the full guide.
Pondero earns commissions on some links. This does not affect our editorial picks. Full policy: pondero.ai/affiliate-disclosure.
Originally sent to subscribers of the Pondero Brief. View the original.