Vercel's Dry-Run Deploy Manifest Turns Agent Releases Into a Preflight Lane

A deployment mistake usually shows up too late.

The wrong file ships. The expected file doesn't. A huge asset slips through. A framework is detected differently than the team assumed. Then everyone acts surprised even though the problem was sitting in the package all along.

Vercel's new vercel deploy --dry matters because it gives that package a visible shape before anything uploads.

The interesting part is the manifest, not the flag

Butler has already watched Vercel push more of its operational surface into CLI- and operator-friendly lanes, from web analytics feedback loops to scriptable Flags targeting to Agent Runs moving into MCP and CLI.

The dry-run deployment feature extends that pattern.

Vercel says teams can now run vercel deploy --dry from a linked project to preview the detected framework preset plus which files are included or ignored. With --format=json, the command returns a machine-readable manifest including path inclusion, ignored paths, largest files, file modes, hashes, and directory size distribution.

That is not just a convenience for humans glancing at terminal output.

It is a control point for automation.

Why this matters for agent-driven releases

Agents are getting better at preparing release candidates. They are still very capable of packaging the wrong thing.

Sometimes the code is fine and the failure lives in the deploy boundary:

• the wrong file is ignored
• the right file is missing
• a build artifact bloats the upload
• framework detection changes unexpectedly
• a permission or mode bit looks strange

Before a feature like this, teams often learned that truth only after a deployment was created or uploaded.

A dry-run manifest inserts a checkpoint before the side effect.

That matters because it creates a reversible inspect-and-fix loop. The agent can run the preflight, compare the manifest against expectations, repair .vercelignore or packaging issues, and rerun until the output looks sane.

JSON output is the real unlock

The human-readable summary is nice.

The bigger operational shift is JSON.

Once the manifest is machine-readable, it becomes easy to gate deploys on concrete expectations:

• expected framework must match
• certain files must be present
• certain paths must be absent
• largest-file thresholds must stay under a limit
• suspicious file modes must fail the preflight

That turns deployment packaging from a vague terminal moment into policy-checkable data.

This is also a cleaner way to debug .vercelignore

Ignore-file bugs are some of the most boring deployment failures, which is exactly why they waste so much time.

Nobody wants to find out after upload that a critical directory was excluded or a junk artifact was included. A dry run gives teams a faster loop for that class of mistake.

For agent operators, this is especially useful because the manifest can become part of the approval packet. Instead of saying the deploy should be fine, the workflow can show exactly what would ship.

What this does not replace

It does not replace staging.

It does not prove runtime behavior.

It does not guarantee the app is correct once deployed.

The value is narrower and more practical: you can now inspect the package before remote creation.

That alone closes a very real gap.

What teams should standardize next

• Add dry-run manifest checks before automated deploy creation.
• Define a few hard expectations: framework, required paths, banned paths, size ceilings.
• Treat .vercelignore and packaging rules as policy, not cleanup trivia.

Vercel's dry-run deploy feature matters because many release failures are packaging failures wearing a different costume.

If an agent can inspect the shipping manifest before upload, you catch a whole class of mistakes while the fix is still cheap.

Related coverage

• Vercel's recent Agent Runs operator surface
• the earlier scriptable targeting surface
• the CLI analytics feedback loop Butler already tracked

AI Disclosure

This article was researched and drafted with AI assistance, then reviewed and edited for clarity, accuracy, and editorial quality.