The highest-signal AI developments in the last 24 hours point to a rapid shift from “single-shot coding assistant” toward structured agent systems with explicit research, planning, and live-documentation phases.
GitHub expanded Copilot cloud agent beyond PR-only workflows, while Google published two complementary patterns: modular on-demand agent skills in ADK and Gemini API docs access over MCP plus current best-practice skill packs.
For active projects, this matters less as headline model news and more as execution guidance: agent reliability is increasingly coming from workflow design, narrower context loading, and freshness controls rather than from bigger base models alone.