Most agent work vanishes into memory. Artifact execution fixes it.

Your agent gets a task. It starts working. Fifteen minutes later, you check in — it’s still "making progress" but has nothing to show. You wait another ten minutes. Still nothing concrete.

This is the vanishing work problem. Your agent is thinking, planning, researching, "making progress" — but producing zero artifacts you can actually see, test, or build on.

I watched our coding agent spend 40 minutes "refactoring the authentication system" with nothing to show for it. No files changed. No commits. No diffs. Just a long explanation of what it was thinking about doing.

The fix is simple: artifact-first execution. Every agent action must produce a concrete deliverable before it can continue.

Rule: If your agent can’t show you what it built in the last 10 minutes, it’s not building — it’s daydreaming.

Here’s the pattern that stops vanishing work:

ARTIFACT_RULE = """
Every task step must produce a deliverable:
- Code changes → commit with clear message
- Research → markdown file with findings
- Analysis → structured document with conclusions
- Planning → written spec with next steps

No "I’m working on it" updates without artifacts.
No "almost done" without something to review.
No "making progress" without measurable output.
"""

We added this to our agent’s system prompt and immediately saw the difference. Instead of long thinking sessions, we get:

• Incremental commits every 5-10 minutes
• Draft documents that get refined, not rewritten
• Working prototypes instead of perfect plans
• Reviewable progress at every checkpoint

The key insight: thinking time doesn’t compound, but artifacts do. A half-finished function is infinitely more valuable than a perfect mental model of the function.

Your agent should work like a craftsperson, not a philosopher. Craftspeople always have something to show — even if it’s rough, even if it’s incomplete. Philosophers can think for hours and produce nothing tangible.

Set up artifact checkpoints every 10 minutes. If your agent can’t point to a file, a commit, a document, or a concrete output, it’s not working — it’s procrastinating.

The agents that ship are the ones that build in public, even to an audience of one.