Fixing a PR resolves the problem in front of you. Lauren Tan’s approach to AI coding asks one more question: if the next PR makes the same mistake, who will catch it first? If the answer is still the reviewer, someone has to make the same judgment again. Put that judgment in an interface or a check, and the agent can get feedback before asking for review.
Tan, also known as Poteto, is a principal engineer on SpaceXSI’s Grok Bot team. She described how the team builds these constraints in All Code Is Slop (Except Mine), given at AI Engineer CODE on November 11, 2026. A data import failure shows what it takes to fix a patch today and prevent the same mistake tomorrow.
1. Set the criteria: distinguish a retry from a new job
Suppose an import writes its records, but the client times out before receiving confirmation and retries. Decide what should happen in two different cases:
- Retry the same job: the records it already wrote appear only once.
- Create a new job with the same file: the new job can still run.
The file is identical in both cases; the intended behavior differs. Tell the agent only to “prevent duplicate imports” and it might deduplicate by file, blocking legitimate new jobs. Settle that distinction before asking for a fix.
2. Fix this patch: make the same check fail before and pass after
In a test environment, reproduce the case where the records were written but the confirmation never arrived. Retry the original job and inspect the stored records. Have the agent demonstrate the duplicate writes on the old version, then run the same check after the fix to show that the records are written only once. Start a new job with the same file to confirm that it still writes its records.
The reviewer can now see that the check distinguishes the bug from the intended behavior. A passing run on the new version alone doesn’t show that the check ever reached the original failure. Checking only for duplicates could even accept a patch that rejects every subsequent import. Set the acceptance criteria before the patch; have the agent produce evidence by running the program.
3. Protect future changes: run the regression check before merge
Keep the check and run it automatically before merge. When the same bug returns, the agent gets a failure it can act on before a reviewer has to explain it again.
A test only protects the path it exercises. Add an import entry point that bypasses the old protection, and the existing check might still pass. If several entry points must obey the same rule, enforce it in the interface they all go through. Migrate callers off the old path so it doesn’t become a template for future patches.
Here, the regression test detects duplicate writes, and the shared interface makes new entry points follow the same rule. AGENTS.md can explain why imports are distinguished by job; the code enforces that decision.
---
config:
flowchart:
curve: linear
rankSpacing: 28
---
flowchart TB
accTitle: Fix this patch and protect the next one
accDescr: The solid blue arrow sends a failed check back to the current patch. The dashed amber arrows show how a recurring review finding changes the constraints on future patches.
H["Set acceptance criteria"] --> A["Agent edits code"]
A --> V["Run acceptance checks"]
V -->|Fail: revise patch| A
V -->|Pass: share evidence| D["Review the change"]
D -->|Find a recurring issue| F["Update APIs or checks"]
F -->|Protect the next patch| A
classDef verify fill:#e7f0ff,stroke:#2563eb,color:#132b55,stroke-width:2px
classDef feedback fill:#fff1d6,stroke:#b45309,color:#713508,stroke-width:2px
classDef review fill:#eef4f8,stroke:#58708a,color:#20374f
class A,V verify
class F feedback
class H,D review
linkStyle default stroke:#58708a,color:#334b65,stroke-width:1.5px
linkStyle 2 stroke:#2563eb,color:#1e40af,stroke-width:2.5px
linkStyle 4,5 stroke:#b45309,color:#92400e,stroke-width:2.5px,stroke-dasharray:5 3
Start with a correction that keeps coming up in review. Can you make the next patch that repeats the mistake fail before review? Get that feedback working for one agent before adding parallelism. Otherwise, you’ll have more patches waiting for the same human check.
Further reading: Anthropic’s multi-agent harness covers how to split roles when a task calls for separate generation and evaluation.
Sources:
- All Code Is Slop (Except Mine): talk recording (Lauren Tan, AI Engineer CODE, talk given November 11, 2026)
- All Code Is Slop (Except Mine): conference write-up and transcript (AI Engineer, November 11, 2026 conference)
- All Code Is Slop (Except Mine): slide index and examples (AI Engineer, 2026 conference materials)
Author: Cyber Herald
Original Author: Lauren Tan
Original Source: https://www.youtube.com/watch?v=O_zcc6Xl12g
Adaptation Date: 2026-11-19
Last Updated: 2026-11-19