What is shadow mode testing for AI agents?

Shadow mode testing means running an AI agent on real requests while its writes are held back, then comparing what it would have done with what people actually did. It shows whether the agent makes the right decisions on real inputs. It can't show what happens when the writes land: duplicates, partial writes, false "done" reports.

Variants

  • Pure shadow. The agent proposes; nothing is written anywhere.
  • Draft mode. The agent writes drafts or on-hold documents, and a person posts or discards them.
  • Shadow for a new version. The current version keeps working while the new one runs alongside on the same requests.

What it shows and what it misses

Shadow mode is good at decision quality: did the agent pick the right supplier, the right quantity, the right matter? It can't show what happens when the write actually lands, because it never does. Duplicate writes on retry, partial writes, hallucinated success, approval limits and effects on linked documents only show up when the write runs.

Using it with release gates

Use shadow mode for decision quality on real inputs. Test the writes separately in a test copy of each system before go-live, with injected failures.

On the blog

AI agent evals vs testing · Fault injection for AI agents · Vertical AI agent · All terms

FAQ

For decision quality, it is a good start. It can't test what happens when writes land, such as duplicates on retry or partial writes, so test those separately.

In pure shadow mode nothing is written. In draft mode the agent writes drafts or on-hold documents that a person posts or discards.

Long enough to see the cases that matter in your process, including the rare ones. Compare field by field, not just "looks right".

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.

Start free with Community Edition

20 runs a month, no credit card.

Your agent writes into ERP, CRM or billing for customers? See Enterprise.