the dam brief #15

The Headline

the week's biggest story

Anthropic launches Claude Sonnet 5 for cheaper agentic runs

Anthropic launched Claude Sonnet 5 on June 30, calling it its most agentic Sonnet model yet. the company says it can make plans, use browsers and terminals, and run autonomously at a level that recently required larger and more expensive models.

the release landed alongside Claude Science, a research workbench for scientists and pharma teams, and a July 1 return for Fable 5 after US export controls were lifted.

(Source)

The Build

tools, models, and dev releases

quick signal before the rabbit hole: these six repos add up to 105k+ GitHub stars. the useful part is what each one lets you try this weekend.

• Strix stress-tests apps with AI penetration testing - a runnable path for finding and fixing app vulnerabilities before a human red team gets expensive (GitHub)

• OmniRoute gives coding agents one gateway across model providers - a TypeScript gateway with MCP, A2A, and provider fallback if your agent stack is glued together by API keys (GitHub)

• video-use edits video with coding agents - a browser-use project with a small enough README to try if your content pipeline still treats video as manual timeline work (GitHub)

• cognee gives agents self-hosted long-term memory - a knowledge-graph memory platform with a concrete install path for teams tired of agents forgetting the last run (GitHub)

• herdr multiplexes terminal agents - a Rust terminal pattern for supervising multiple agent runs without turning your desktop into a window graveyard (GitHub)

• CubeSandbox hardens agent execution - TencentCloud's Rust sandbox for instant, concurrent, lightweight isolation around agent code (GitHub)

Founder Fuel

funding, acquisitions, and market moves

• Microsoft launches an AI deployment company with a $2.5B commitment - TechCrunch reports that the new group follows Amazon, OpenAI, and Anthropic into the business of taking AI from models into enterprise operations (More)

• OpenAI reportedly proposes giving 5% of its equity to a US sovereign wealth fund - the proposal would turn private frontier-lab upside into a public-finance question instead of a normal cap-table story (More)

• Anthropic discusses a custom chip with Samsung - the talks add another frontier lab to the custom-silicon race after OpenAI's Broadcom partnership moved compute strategy closer to the model roadmap (More)

• Bending Spoons jumps 40% on its first trading day - the Milan-based software buyer went public after building a portfolio around brands including Evernote, Meetup, Vimeo, AOL, and Eventbrite (More)

Longevity & Health

aging, biotech, and health research

• OpenAI releases GeneBench-Pro for genomics and biology evaluation - the benchmark tests AI systems on complex real-world datasets in genomics, biology, and scientific research rather than generic multiple-choice biology prompts (More)

• Anthropic ships Claude Science for research and pharma teams - the workbench integrates common scientific tools, produces audit trails, and targets repeatable research workflows rather than one-off chat answers (More)

• scientists trace a possible route for Alzheimer's spread through the brain - ScienceDaily reports that blocking toxic Tau-carrying packages before they reach healthy cells reduced a suspected disease-spread mechanism in lab work (More)

The World

regulation, policy, and geopolitics

• Google must pay a €4.1B Android fine in Europe - the BBC reports that the court rejected Google's challenge over using Android to block rivals, keeping platform bundling inside Europe's competition fight (More)

• UK police warn parents about public child photos and AI abuse - the National Crime Agency says ordinary images can be repurposed into child abuse material, turning family sharing into a live safety and platform-governance issue (More)

• AWS publishes its frontier-model release process for Bedrock - the post details security, governance, and customer-release checks for new frontier models as cloud providers become gatekeepers for model access (More)

Go Deeper

long reads, podcasts, and documentaries

• standard benchmarks underestimate what agents can do - The Decoder | staff. the UK's AI Security Institute found that capped compute budgets hide agent capability, especially on software engineering tasks. (Read)

• Google DeepMind unionization talks are off to a rocky start - Wired | Paresh Dave. a useful labor-side view of what happens when frontier AI work becomes normal workplace politics. (Read)

• paradise revisited - Longreads | staff. a Galapagos and Darwin piece for anyone who needs one long read that is not another model benchmark. (Read)

• constructions, concoctions, conjurations - Longreads | staff. a reading list on constructed languages, useful if you like watching humans build protocols without calling them protocols. (Read)

The Colony

events and community

• next friday: July 10 - build fridays Hamburg - bring a laptop, pick one thing, and make visible progress with other builders in the room (RSVP)

• next sunday: July 12 - agentic coding studio, Hamburg - a focused session for builders working with coding agents, hosted by AI BEAVERS with Malte Sussdorff and Rodrigo R. Espitia (RSVP)

• new for members: the community perks hub is live with Notion, Linear, and SAGEOBOT benefits for eligible AI BEAVERS members after newsletter confirmation and profile completion (Open perks)