What are partial writes?
A partial write is when an AI agent completes some steps of a multi-step task and then stops, leaving a half-finished change in a business system: an order without its lines, a refund without its credit note, a contact moved without its deals. The right result is the full change or a clean, reported state.
Why it happens
Multi-step tasks fail in the middle: a server error on step three, a timeout, a validation error on one line item. If the agent has no rollback or resume logic, the first steps stay saved. If it then reports success, the partial write becomes hallucinated success as well.
How to test it
Run each multi-step task in a test environment and force a failure at each step in turn, using fault injection. After each run, check the system. Pass: either the complete change, or a clean state that the agent reported clearly and that someone can safely finish or undo. Fail: a half-written record, especially one the agent called "done".
On the blog
- AI agent partial failures and rollback testing
- AI agent hallucinated success
- Duplicate writes and idempotency testing
Related terms
Hallucinated success · Fault injection for AI agents · Duplicate writes · All terms
FAQ
Either the full change or a clean, reported state that someone can safely finish or undo. Never a half-written record presented as complete.
Force a failure at each step of each multi-step task in a test environment, then check what the system holds.
No. A partial write leaves too little behind; a duplicate write leaves too much. Both are triggered by failures in the middle of a task, so test them together.
Try it on your own agent
Describe a task in plain words. Agentiqa does it step by step and tells you if it worked. Every failure comes with step-by-step reproduction.
20 runs a month, no credit card.
Your agent writes into ERP, CRM or billing for customers? See Enterprise.