Self-reflection

Pipelines get external validators. Agents can do something pipelines can't: check their own work mid-task and change course. Used precisely, reflection converts a class of failures into recoveries. Used naively, it's expensive self-congratulation. Three patterns earn their tokens:

• Verify-after-act: after a consequential write, read it back. Updated the ticket? get_ticket and confirm the note landed as intended. Trivial, mechanical, catches silent API half-failures that would otherwise surface as customer complaints. Cheapest reflection there is — make it a rule for every write tool.
• Pre-finalize review: before declaring done, one structured pass — 'Re-read the original ticket. List each thing the customer asked for and where your resolution addressed it. Anything unaddressed → not done.' This catches the classic agent failure of resolving a task instead of the task (answering the damage claim, missing the 'also, change my address' buried in paragraph two).
• Retry-with-lessons: on a failed approach, force an explicit lesson before the next attempt — 'state what you tried, why it failed, and what you'll do differently.' Without the articulation step, agents retry the same move with cosmetic variations; with it, retries actually branch. (Your validator retry-with-feedback pattern, running inside the loop.)

The limits — respect them

• Reflection shares the blind spots of the reasoning it reviews. An agent that misread the policy will 'verify' its refund against the same misreading. Self-review catches slips (skipped steps, half-done writes, unmet sub-requests) far better than misconceptions. For misconceptions you need something outside the loop: a differently-prompted checker (your judge patterns), a validator, a human.
• Self-assessed confidence inflates. 'Are you sure?' produces yes. Structured checklists ('list each request → point to its resolution') work because they demand evidence, not vibes — the same lesson as judges needing anchored rubrics.
• Budget it. Reflection multiplies turns. Verify-after-act on writes: always. Pre-finalize review: once. Open-ended 'reflect until satisfied': never — that's how 40-turn trajectories are born.