Reproduction First

Problem

You report a bug: "Users occasionally get logged out mid-session." The reads the auth code and proposes a fix — extends the refresh window. You deploy. It still happens. The agent proposes another fix — adds a retry on the refresh call. Still happens. A third: clock-skew tolerance. Each fix is plausible, each is wrong, and you are now three deploys into a bug that is exactly as alive as when you started.

Every one of those fixes was a guess. The agent never reproduced the logout. It pattern-matched "intermittent auth failure" to common causes and proposed the textbook remedies for each. With no way to observe the actual failure, it could not tell which cause was yours, so it cycled through the catalog. Plausible-but-unverified is the defining texture of guess-driven debugging, and agents are extremely fluent at producing it.

The loop is self-sustaining. Each failed fix changes the code, which changes what the next guess reasons about, which justifies the next plausible fix. You can stay in this loop indefinitely because it always produces a next thing to try. What it never produces is a fix that works, because nothing in the loop ever confronts the actual bug.

Solution

Stop asking for fixes. Make the next deliverable a deterministic reproduction — a command, script, or failing test that triggers the bug on demand. No fix attempt is allowed until the bug can be summoned at will.

Change what you are asking for:

Stop proposing fixes. The logout bug is not reproduced yet, so
every fix is a guess.
 
Next deliverable: a failing test (or a script) that reproduces
the mid-session logout deterministically. It must fail for the
real reason, not a mocked stand-in. Do not change application
code in this step. If you cannot reproduce it, report what
you need — logs, repro steps, the conditions — instead of
guessing again.

Reproduction is a step with its own gate. Treat "can we trigger it on command?" as a hard checkpoint. Until it is green, debugging has not actually started:

Phase	The question	Exit condition
Reproduce	Can we trigger it on demand?	A command that fails every time
Diagnose	What does the repro reveal?	The mechanism, named from evidence
Fix	Does the change flip the repro?	Was red, now green
Guard	Can it silently come back?	Repro kept as a permanent test

A real reproduction needs the real cause. A test that mocks the suspected failure "reproduces" your assumption, not the bug — it will pass after a fix and the bug will still ship. The reproduction has to fail for the actual reason, which is exactly why producing it forces the understanding that guessing skips.

Keep the reproduction after the fix. The artifact that proves the bug exists is, unchanged, the artifact that proves it never comes back. The reproduction graduates directly into a Regression Guard — debugging output becomes permanent protection at zero extra cost.

A Confident Fix Without a Reproduction Is Still a Guess

The danger is not that the agent refuses to help — it is that it produces a fluent, well-reasoned, entirely unverified fix and you, several failures deep and wanting to be done, accept it. Confidence in the explanation is not evidence about the bug. Hold the line: no reproduction, no fix. "I cannot reproduce it yet, here is what I need" is real progress. The fourth plausible patch is not.

Signals

Multiple deployed "fixes" and the bug is still exactly as alive
Fixes are justified by plausible theory, never by an observed failure
Each attempt targets a different root cause than the last
Nobody can state the command that makes the bug happen
The debugging session generates endless next-things-to-try and no resolution

Consequences

Benefits:

Converts guessing into diagnosis — fixes confront the bug, not a theory
A fix's success becomes verifiable: the repro flips red to green
The reproduction graduates into a permanent regression test for free
Ends the infinite plausible-fix loop with a hard, objective gate
Forces real understanding, because a true repro can't be faked into existence

Costs:

Some bugs are genuinely hard to reproduce — heisenbugs, races, prod-only state
Reproduction-first feels slower right up until it is the only thing that worked
A mocked or approximate repro gives false confidence — it must fail for the real reason
Requires holding the line against an agent (and yourself) wanting to just try one more fix

Relationship Map