A burst of Copilot releases — desktop app, REST-triggered cloud agents, IDE sessions and a new enterprise base model — shows where agentic development is heading: isolated tasks, auditable runs, and pull requests as the point of control.
Read the piece →
Anthropic says Opus 4.7 improves on Opus 4.6 for difficult software engineering tasks, adds an xhigh effort level, launches task budgets in public beta, and brings an /ultrareview command to Claude Code.
For agencies building AI over client documents, Google’s update tackles three recurring RAG problems: visual assets, noisy corpora and answers that cannot be traced back to a source page.
Laravel’s AI SDK has gained sub-agent support, allowing a parent agent to delegate a task to another agent with its own prompt, tools, model provider and runtime configuration.
Cloudflare has moved Browser Run onto Cloudflare Containers, lifting Workers binding limits to 60 browser starts per minute and 120 concurrent browsers while cutting Quick Action response times by more than half.
AI systems that design proteins and biological agents are already useful for drug discovery and basic science. The uncomfortable question is how to preserve that value while reducing misuse risk.
Mars missions already rely on orbiters as data relays. NASA’s new RFP points to something more deliberate: a service-like communications layer at the Red Planet, with industry asked to show readiness by 2030.
NASA Earth Observatory’s new view of artificial light at night shows that the planet is not simply getting brighter in a smooth line. Using Black Marble data from VIIRS instruments on Suomi-NPP, NOAA-20 and NOAA-21, researchers mapped where night-time light intensified and where it dimmed between 2014 and 2022. Gold areas mark brightening; purple areas mark dimming. The pattern reflects construction, industrial booms and busts, blackouts, policy-driven lighting changes and other human activity. It is a reminder that satellite images of Earth at night are not just beautiful: they are a proxy for energy use, urban change, conflict, recovery and the way modern life leaks into the dark.
On 18 May 1969, Apollo 10 lifted off from Kennedy Space Center on the final full rehearsal before the first Moon landing. NASA described the mission as a dress rehearsal: Thomas Stafford, John Young and Eugene Cernan took the command and lunar modules to lunar orbit, tested procedures, communications and rendezvous, and flew the lunar module Snoopy down to within about 15 kilometres of the Moon’s surface without landing. The mission mattered because Apollo 11 depended on more than courage and a rocket. It needed a validated operational stack: navigation, staging, crew choreography, ground control and abort paths. Apollo 10 turned the landing plan from theory into practised procedure.
Most coding-agent benchmarks still look like small bug-fix tasks: one issue, one repository, pass or fail. RoadmapBench asks a more useful question for real software teams: can an agent handle longer-horizon work across version upgrades, where changes span many files and require coordinated progress rather than a single patch? The authors introduce 115 tasks grounded in 15 real-world repositories, with evaluation aimed at multi-target development rather than only unit-test success. For anyone using agents in production codebases, the paper is a reminder that the frontier is not just writing correct snippets. It is planning, sequencing, preserving architecture and surviving the messy middle of an upgrade.
arXiv:2605.15846 →