9 Critical Failure Patterns of Coding Agents
Columbia University DAPLab (blog post by Reya Vir) · v2026-01 · first published 2026-01-08
Scope
Nine failure patterns that the authors saw while building 15+ applications with five coding agents and vibe-coding tools (Claude, Cline, Cursor, v0, Replit; testing as of November 2025). The post says it documented "hundreds of failures" during development and grouped them into the nine patterns. It is a blog post about building web applications, not a peer-reviewed study, and it gives no counts per pattern.
Why it's ingested
A first-hand account of where coding agents fail when building complete applications. Its patterns are about application quality (UI, state, business rules, data model, integrations), so only some of them have a counterpart among the SAICA failure modes; data/crosswalks/daplab-to-saica.yml lists which and gives a reason for each one that has none.
Categories & crosswalks
Each row is one category of this taxonomy with the SAICA failure modes mapped to it and how close each mapping is: “exact” means the same failure, “partial” an overlap, “related” a neighbouring category. 7 mappings cover 6 of the 9 categories; the 3 with no mapping are listed under them. The category ids and labels were last compared with the source on 2026-11-21.
Mapping drafted by an AI model from the source text; not yet reviewed by a person.
-
DAP-03Business Logic MismatchAgents "often fail to correctly implement" the rules, permissions and calculations of the use case: "code that runs without errors, but it produces the wrong output because the underlying logic is flawed." Example: a 10% discount applied per item instead of to the whole cart.
SAICA-KG crosswalks: -
DAP-05API & External Service Integration FailuresAgents "frequently fail to call APIs correctly". The agent may prefer code that appears to change the UI over code that works, and "may hallucinate fake inputs like environment variables or API keys instead of asking for the real values". Example: a feature built on a placeholder key, so the call fails and a hard-coded response is shown.
SAICA-KG crosswalks: -
DAP-06Security Vulnerabilities"Agents often introduce security vulnerabilities – even basic ones." They lack an understanding of data sensitivity and access control and "frequently expose sensitive data like API keys" or get access control wrong. Example: answers from a private admin codebase given to regular users.
SAICA-KG crosswalks: -
DAP-07Repeated Code"Agents often write repeated code, adding duplicate functions with very similar logic rather than abstracting it." Example: the same API set-up code copied into two fetch functions instead of a shared helper.
SAICA-KG crosswalks: -
DAP-08Codebase Awareness & Refactoring Issues"Agents lose context in larger projects": as files are added the agent loses track of the architecture and "may mistakenly edit the wrong component or introduce breaking changes". As a by-product, agents "often re-implement an algorithm from scratch, rather than using a well-established software package". Example: a hand-written Levenshtein distance instead of importing the editdistance library.
SAICA-KG crosswalks: -
DAP-09Exception & Error Handling"The agent prioritizes execution over correctness": it adds "just enough error handling so that the app runs end-to-end without crashing, but fails to inform the user when things go wrong". Example: a caught error logged only to the developer console while the user sees a stuck "Generating…" message.
SAICA-KG crosswalks:
Categories with no SAICA failure mode
These 3 categories have no mapping. The reason is the one the crosswalk file states; where it states none, that is said.
-
DAP-01Presentation & UI Grounding Mismatch"Agents struggle to translate visual requests into code." They cannot see the interface, so requests such as moving a button, resizing a component or aligning items often give layouts that do not match what the user asked for. Example in the post: a calendar that should be a grid is shown as one vertical line.
Why it is not mapped: Presentation & UI Grounding Mismatch: no SAICA mode covers a layout that does not match a visual request.
-
DAP-02State Management FailuresThe agent fails to keep the application's in-memory state correct (cart contents, list order, the current step of a flow), "often fails to manage shared state between components when refactoring" and struggles to decide what state to store. Example: a drag-and-drop reorder that updates only part of the list.
Why it is not mapped: State Management Failures: defined by the part of the application (in-memory state), not by a SAICA failure definition. One instance would usually be labelled logic_error, SAICA's residual class; mapping the whole pattern there would add no information.
-
DAP-04Data Management ErrorsAgents "struggle with data logic and fail to understand the data model (even one it generated itself!)". The lack of schema awareness leads to redundant columns or a schema that is not updated when new prompts arrive. Example: passing the wrong record id to an update.
Why it is not mapped: Data Management Errors: schema design and data-model awareness. Same reason as DAP-02.
Notes
Rebuilt on 2026-11-21 from the live post (HTTP 200, published 2026-01-08) and checked against the earliest Wayback Machine capture (2026-01-17), which lists the same nine headings. Labels are the post's own headings; quoted text in the descriptions is copied from the post. The ids DAP-01 to DAP-09 are SAICA-KG's, formed from the post's own numbering 1 to 9; the post publishes no identifiers. Retraction. Until 2026-11-21 this file held nine labels that are not in the post: Incomplete Task Execution, Fabricated References, Stale API Usage, Cascading Error Recovery, Context Window Overflow, Tool Misinvocation, Silent Test-Passing Workaround, Destructive Action and Goal Drift. The file's own note called them "paraphrased for seed" and asked for a check that was not done. The post does not describe agents editing tests, running destructive commands, claiming completion with work undone or fix attempts that add new errors. The same ids therefore now name different, real patterns; anything that cited an old DAP id by meaning must be re-read. The correction was first made on 2026-08-18 (commit 00422e2) and was lost when the branch was reset on 2026-11-19. Five incident nodes built on the invented labels were retracted; see data/TOMBSTONES.yml and archive/retracted-incidents/README.md.