Retry Is Not Recovery

Watch what happens after the first failure

Every complex system fails.

The important question is what the system does after the failure becomes visible.

Does it recognize the problem? Does it understand why the path failed? Does it preserve useful partial work? Does it choose a meaningfully different strategy? Does it know when the remaining constraint requires a person?

Retrying the same reasoning is weak recovery

A 2024 ICLR evaluation tested intrinsic self-correction across multiple reasoning tasks and found that models often failed to improve and sometimes degraded when asked to correct themselves without reliable external feedback. The study reports task- and model-specific results rather than one honest universal degradation percentage, so we should not invent a headline number here.

That result is bounded to the tested models and reasoning tasks. It does not mean AI cannot recover.

Asking the same model to think again can be useful, but by itself it does not establish recovery.

A recovery sequence

We think useful recovery has at least six steps:

1. Detect — recognize that the expected state was not reached.
2. Classify — identify whether the failure came from reasoning, tools, state, permissions, dependencies, or the plan itself.
3. Preserve — keep the parts of the work that are still valid.
4. Replan — choose a materially different safe path when the original path is no longer justified.
5. Verify — confirm that the new path actually restored a usable state.
6. Escalate — involve a human only when the remaining decision, authority, or constraint genuinely belongs to them.

The easiest recovery failure to spot

One simple signal of bad recovery is repeated-path behavior.

If the system encounters the same blocker, uses the same method, and returns the same failure, it has not learned anything operationally.

A trustworthy system should be able to say: this route failed for a specific reason, so I am changing the route—or I am stopping because no safe route remains.

When recovery is weak, the user becomes incident response

When recovery is weak, the human becomes the incident-response system.

They diagnose the failure, reconstruct state, repeat instructions, clean up partial changes, and decide where to restart.

That is not only a technical defect. It is a Human–AI Experience defect.

Do not call it self-healing until it can prove recovery

“Self-healing” is easy language to use and hard behavior to prove.

A system that retries automatically is not necessarily self-healing.

A stronger claim would require evidence that it can recognize classes of failure, preserve constraints, restore or reach a safe state, avoid repeated damage, and reduce recurrence over time.

After a failed attempt

When an AI fails, do not ask only “Can you try again?”

Ask: “Why did that path fail, what are you changing, and how will we know the new path worked?”

If the system cannot explain why the first path failed or what changed in the next attempt, treat the next attempt as another retry.