Anthropic teaches Claude agents to “dream” between sessions

simonwillison.net · 2026-05-06

At Code w/ Claude, Anthropic launched three Managed-Agents features: multi-agent orchestration and “outcomes” in public beta, plus a research-preview Dreaming mode that lets agents review old sessions overnight and rewrite their own memory. Harvey says completion rates roughly 6× in early tests.

Anthropic ships ten agent templates for Wall Street workflows

anthropic.com · 2026-05-05

The templates target the busywork of finance — building pitchbooks, screening KYC files, closing the books at month-end — and ship as plugins for Claude Cowork and Claude Code, plus cookbooks for Managed Agents.

Microsoft re-frames Copilot as an “agent operating layer” for work

microsoft.com · 2026-05-05

Copilot Cowork moves from chat assistant to a fabric where Copilot delegates multi-step work to agents that touch Outlook, Word, Excel and PowerPoint. The pitch: skills, integrations and devices stitched into one place where you brief once and get artifacts back.

xAI ships a low-latency voice agent API and standalone speech models

docs.x.ai · 2026-05-04

grok-voice-think-fast-1.0 is xAI’s first flagship voice-agent model, tuned for tool use, 25+ languages, and customer-support workloads. Speech-to-text and text-to-speech were broken out as separate APIs with diarization and timestamps.

Hugging Face cuts a major version of Transformers

huggingface.co · 2026-05-06

Transformers v5 simplifies the model-definition pattern that most of the open-source AI ecosystem still imports. It’s a once-every-few-years cleanup, and it lands alongside an updated huggingface_hub client the same week.

Kimi-maker Moonshot AI hits a $20B valuation in a Meituan-led round (unverified)

bloomberg.com · 2026-05-07

Reports put the new round at roughly $2B and a $20B post-money — the clearest sign yet that China’s consumer-AI race has a bona-fide private-market champion. Bloomberg paywall blocked verification of the headline.

Anthropic says API traffic is up 17× year-over-year

simonwillison.net · 2026-05-06

A throwaway slide at Code w/ Claude shows API volume on the Claude platform 17× higher than a year ago. It is the firmest public usage number Anthropic has shared in a while and pairs with a fresh SpaceX compute deal.

Anthropic teams with Blackstone, H&F and Goldman on an enterprise services co

anthropic.com · 2026-05-04

A new joint venture will package Claude into deployable AI services for Fortune-500 buyers. The interesting part: the money and distribution come from three of the most relationship-heavy firms in finance, not from another tech company.

Perplexity’s Comet browser is now on every major platform

perplexity.ai · 2026-05-05

Comet rolled out to all iOS users this week, finishing the set after Mac, Windows and Android. The same release added MCP server connections for Pro / Max / Enterprise plans, turning Perplexity into a usable agent harness for outside data.

Claude Code 2.1.133: subagents finally see project skills

docs.claude.com · 2026-05-07

Three releases this week. Headline fixes: subagents discover project / user / plugin skills, hooks receive an effort.level input, and a 1-hour prompt-cache TTL silently downgrading to 5 minutes is gone. New --plugin-url flag fetches plugin zips from URLs.

npm i -g @anthropic-ai/claude-code@2.1.133

#release #agent

Gemini Embedding 2 hits GA on the Gemini API and Vertex AI

blog.google · 2026-05-06

Google’s second-gen embedding model is out of preview. Same multilingual setup, longer input window, and the same text-embedding-gemini-2 handle on both surfaces — easy drop-in if you were on v1.

DeepSeek-V4 lands on Hugging Face with mHC and a hybrid attention stack

huggingface.co · 2026-05-05

V4 swaps Multi-head Latent Attention for a hybrid local + long-range design and replaces residual connections with manifold-constrained “hyper-connections”. Flash, Pro and -Base variants are all up.

Twill.ai (YC S25) — delegate to cloud agents, get back PRs

twill.ai · 2026-05-06

A managed harness that runs a Codex/Claude-Code-style agent in the cloud against your repo and opens PRs back. Aimed at the boring half of the backlog: dependency bumps, lint debt, repetitive refactors.

Freestyle launches sandboxes purpose-built for coding agents

freestyle.sh · 2026-05-05

A sandbox runtime for letting agents run arbitrary code without trashing the host. Sub-second cold start, ephemeral filesystems, and a network policy you can describe in JSON — useful as a primitive if you’re building or evaluating a coding agent.

Kampala (YC W26) — reverse-engineer mobile and web apps into APIs

zatanna.ai · 2026-05-07

Point Kampala at an app and it produces a typed API surface plus a small client SDK. The pitch: scrape-bots that don’t rot, by replaying the actual app under instrumentation rather than guessing at endpoints.