the dam brief #11

The Headline

the week's biggest story

agent pull requests are everywhere, and review is becoming the workflow

GitHub published a practical guide on May 7 for reviewing agent-generated pull requests, with advice on scoping changes, checking hidden assumptions, reading tests as evidence, and watching for technical debt that looks productive at first glance.

it sits beside GitHub's May 6 post on validating agentic behavior when the correct answer is not deterministic. together, the pieces make one point operational: serious teams now need review systems for agent work, not just better prompts.

(Source)

The Build

tools, models, and dev releases

• GitHub shows how to validate agentic behavior when correctness is not deterministic - the useful angle is evaluation design: log trajectories, define domination checks, and test outcomes without pretending every agent run has one golden answer (More)

• GitHub adds flexible secrets and variables for Copilot cloud agent - delegated background work can now receive environment-specific configuration through GitHub Actions, so teams should separate agent permissions by repo, task, and deployment target (More)

• GitHub exposes Copilot code review comment types in usage metrics - teams can now see what kinds of review suggestions Copilot generates, which makes agent review measurable instead of another invisible developer-experience line item (More)

• OpenAI ships new voice intelligence features in the API - TechCrunch reports the release targets customer service, education, and creator workflows, so builders should test latency, interruption handling, and transcripts before treating voice as a UI skin (More)

• Mozilla says Anthropic's Mythos found 271 Firefox vulnerabilities with almost no false positives - AI-assisted security work is moving from toy audits into browser-scale bug discovery, which means teams need triage loops and repro evidence before they trust automated findings (More)

Founder Fuel

funding, acquisitions, and market moves

• Moonshot AI raises $2B at a $20B valuation - TechCrunch reports the Chinese lab crossed $200M in annualized recurring revenue in April, driven by paid subscriptions and API usage around its open-source model line (More)

• Ramp is in talks to raise at a $40B-plus valuation - the reported $750M round would come six months after Ramp reached $32B, showing finance workflow software is still being priced as a control surface for company operations (More)

• Cloudflare says AI made 1,100 jobs obsolete while revenue hit a record - CEO Matthew Prince tied the company's first large layoff to support automation, a blunt signal that AI efficiency claims now show up in headcount plans (More)

• Voi founders raise $16M for Stockholm AI startup Pit - Andreessen Horowitz led the seed round, adding another European founder team to the race for AI-native products outside the usual Bay Area gravity well (More)

Longevity & Health

aging, biotech, and health research

• researchers find a way to limit dangerous belly fat after injury in older women - ScienceDaily reports that testosterone gel plus exercise helped prevent visceral-fat gain after hip fracture, pointing at recovery as a longevity intervention window (More)

• boosting one brain-support protein helps mice fight Alzheimer's plaque - raising Sox9 activity in astrocytes reduced plaque buildup and preserved cognition in mouse models, still early but cleaner than another amyloid-only framing (More)

• a common constipation drug shows kidney-protection signals in a clinical trial - a 150-patient study linked lubiprostone to preserved kidney function through gut bacteria and spermidine, which makes the gut-kidney axis less abstract (More)

• lab-grown insulin cells reverse diabetes in mice - Swedish researchers improved stem-cell-derived insulin cells that respond to glucose and restored blood-sugar control after transplantation in diabetic mice (More)

The World

regulation, policy, and geopolitics

• the European Commission reports on the Migration and Asylum Pact rollout - the May 7 update tracks how member states are preparing for new border, screening, asylum, and return rules before the pact applies in 2026 (More)

• Trump gives the EU a July 4 deadline for a trade deal - BBC reports the tariff threat came as a US trade court ruled his global tariff policy violated US law, keeping transatlantic supply-chain planning unstable (More)

• age assurance laws are moving risk down the developer stack - GitHub argues new youth-safety rules increasingly touch operating systems, app stores, package ecosystems, and open-source projects, not just consumer apps (More)

Go Deeper

long reads, podcasts, and documentaries

• what will it take to get AI out of schools? - Longreads | Becca Rothfeld. a sharp counterweight to the assumption that AI-aided education is inevitable, with parents, teachers, and cognitive scientists making the opposite case. (Read)

• software ate my homework - The Atlantic | Ian Bogost. a ransomware attack on Canvas during finals becomes a useful case study in how fragile institutional software dependencies have become. (Read)

• how a congressional primary became an AI proxy battle - The New Yorker | Gideon Lewis-Kraus. one local race turns into a map of OpenAI influence, Anthropic alignment, and the politics of regulating the industry from inside its donor network. (Read)

• do we think too much about the future? - The New Yorker | Joshua Rothman. a clean philosophical reset on why future-thinking can help planning and still distort the present. (Read)

The Colony

events and community

• next sunday: May 10 - Agentic Coding Studio in Hamburg - a smaller room for people building with agents, not just talking about them (RSVP)

• next friday: May 15 - build fridays in Istanbul, Hamburg, and Berlin - three rooms, same rule: bring a laptop and something real to move forward (Istanbul, Hamburg, Berlin)

• friday after next: May 22 - build fridays in Hamburg and London - the colony adds a London room while the Hamburg workbench keeps running (Hamburg, London)