2026-04-04

Signal: Meta's JiT Testing

agents

Signal: Meta’s JiT Testing

Date: 2026-04-04. Source: Meta Engineering blog, Feb 2026.

The claim

Traditional testing frameworks cannot keep pace with agentic development velocity. LLMs should generate ephemeral tests per-change rather than maintaining persistent test suites.

How it works

  1. Developer (or agent) submits a pull request
  2. LLM analyzes the diff
  3. LLM generates targeted tests for that specific change
  4. Tests run, results inform review
  5. Tests are not persisted — they’re disposable

Results

  • 70% reduction in human review load
  • Eliminates test maintenance costs entirely
  • Tests are always relevant (generated from the actual diff)

Why this matters

If agent-generated code becomes the norm, the testing paradigm needs to change. Traditional test suites assume a human developer who understands what they changed and writes tests accordingly. Agent-generated code may change hundreds of files; maintaining a traditional test suite for that output is unsustainable.

JiT testing treats tests as a verification step, not a codebase artifact. That’s a philosophical shift.

Open questions

  • Has this been reproduced outside Meta’s infrastructure?
  • How does it handle integration tests and system-level behavior?
  • What’s the quality floor? Can LLM-generated tests catch subtle bugs?
  • Does this work with local models, or does it require frontier models?

Sources

← all signals