Yo Vercel Community,
New one in the series. 16.3 is still in preview, so this one comes from running 16.3.0-preview.5 in production on our own site since June 30. We run the preview tags on purpose: if something in 16.3 was going to break a production deploy, we’d much rather it broke ours than a client’s.
What’s in it:
- Instant navigations. Every route has to answer “what does the user see immediately”: stream a shell with Suspense, serve cached UI with
'use cache', or opt out withexport const instant = false. Slow navigations become dev-time errors, so users stop being the ones who find them. - Partial prefetching. Twenty chat links in a sidebar used to mean twenty prefetch requests for the same route; now it’s one reusable shell per route, cached for the session.
- Turbopack moves cold compiler state out to the filesystem cache instead of holding it all in RAM, on by default. Vercel’s dashboard app went from 21.5GB to 2GB of dev-server memory.
next devnow detects a coding agent and maintains a block in yourAGENTS.md, the repo has four first-party agent skills, and there’s a/_next/mcpendpoint with acompile_routetool, so an agent can check a route compiles without paying for a full build. Having built our whole content pipeline around agents, we rate this as the sleeper feature of the release.- The edge runtime is deprecated as of this line. Adopting cache components forced our OG image route off it anyway, so the two changes push in the same direction.
- Security releases are monthly from July 20. The first one patches 16.2 and 15.5, which also tells you stable 16.3 is at least a few weeks out.
We built the full instant-navigations adoption on a branch: flags on site-wide, segment configs removed, a 19-test Playwright suite asserting instant navigation, everything green. Then we clicked around the deployments for a week, couldn’t feel the difference, and merged none of it. Our site is fully static, so there was never any server work for a loading shell to hide. Two of the three PRs existed purely so we could click around them and close them. There’s never been a better time to create a couple variations with AI.
Full post has the upgrade notes (one pnpm override gotcha in monorepos), Vercel’s benchmarks, and the current sharp-edges list from the feedback thread. Has anyone flipped cacheComponents on something big yet? The feedback thread is mostly smaller apps so far.
