khawajasafiullah/failures-in-agentic-ai

Failures in Agentic AI

A labelled corpus of 3,089 publicly reported failures of agentic and multi-agent applications, drawn from the issue trackers of 8 orchestration frameworks between 2023 and 2026. Each row is one failure report, coded on five independent dimensions by a deterministic, rule-based instrument, with the text span that decided the primary label kept beside it.

This release accompanies a study of how agentic systems fail, and exists so that the labels, the boundaries and the disagreements can be checked rather than taken on trust.

What is in it

Column group	Contents
Provenance	record_id, source_url, repository, framework, issue_number, created_at
Report context	title, state, issue_labels, comments, reactions, body_chars
Dimension A - failure type	failure_type, failure_type_label, failure_type_confidence, evidence_span
Dimension B - cause class	cause_class, cause_label, cause_secondary, cause_confidence, cause_evidence_grade
Dimensions C, D, E	lifecycle_stage, severity, detection_point
Declared complexity	complexity_tier, complexity_multi_agent, complexity_uses_tools, complexity_uses_memory
Audit trail	matched_rules, taxonomy_version

evidence_span is the sentence or traceback fragment that produced the label; matched_rules lists the rules that fired. Together they let a reader reject any single label without re-running the pipeline. The exact column types are listed in codebook/schema.md.

data/exemplars.parquet holds 49 examples selected to demonstrate each of the 10 failure types, one row per example, with its source URL.

How to load it

from datasets import load_dataset

ds = load_dataset("khawajasafiullah/failures-in-agentic-ai")
df = ds["train"].to_pandas()

df["failure_type_label"].value_counts()
exemplars = load_dataset("khawajasafiullah/failures-in-agentic-ai", data_files="data/exemplars.parquet")

Distribution of failure types

Failure type	Records
Functionality / reasoning defect	748
Compatibility / API-contract defect	561
Data structure / output-parsing defect	422
Tool / integration defect	374
Configuration / environment defect	355
Context / memory / state defect	253
Performance / resource defect	170
Insufficient evidence	92
Agent communication / orchestration defect	90
Safety / guardrail defect	24

Distribution by framework

Framework	Records
AutoGPT	598
LangGraph	539
CrewAI	533
LlamaIndex	436
OpenAI-Agents	337
AutoGen	300
MetaGPT	238
SWE-agent	108

Severity and detection

Severity code	Records
D1_blocking	1,811
D3_degraded	1,053
D2_incorrect_output	174
D4_minor	51

Cause evidence grade	Records
hypothesised	1,545
supported	1,170
demonstrated	374

Across the corpus, 31.9% of failures were never caught by the framework runtime: they reached a user as a wrong result rather than as an error.

Reproducing the labels

The labeller is rule-based and deterministic, so the same input text yields the same label and the same evidence span. The codebook is included as codebook/taxonomy.v1.yaml; it gives every code a definition, an inclusion rule and an explicit exclusion clause. Agreement with a 100-record reference annotation is 51.0% exact (macro-F1 0.54, Cohen's kappa 0.44, moderate); the residual error sits on two codebook boundaries, which are stated in the report rather than tuned away.

Provenance and licensing

• Each row links to its original public report. The full issue text is not redistributed. It belongs to the reporting project, it sometimes contains pasted credentials, and it can be re-fetched from source_url.
• The annotations, the codebook and the evidence quotations are released under CC BY 4.0. Quoted titles and spans remain attributable to their authors.
• No personal data was collected. Author handles were dropped at collection time; what remains is the public issue content needed to justify a label.
• Credential-shaped strings are masked as [redacted: credential] before export, because people paste their own API keys into public issue reports. The stored corpus keeps the original text so it still matches the source; the release does not.
• Collection used only the public GitHub API, rate limited, with every request logged.

Limitations, stated up front

1. This is the reporting population, not all failures. It records what people chose to file on GitHub. Silent failures fixed in-house are absent, so the proportions describe this corpus and not agentic systems in general.
2. Single coder. Labels were produced by one instrument checked against one reference annotation. An independent second-coder pass is prepared but not complete; agreement is therefore reported as moderate, not as validation.
3. Complexity is declared, not instrumented. complexity_tier counts capabilities the report mentions. It under-counts capabilities that were present but undescribed and correlates weakly with report length.
4. Composition drifts over time. A before/after-2025 split shows the mix of failure types changing (Cramer's V = 0.30). Trim by year if you need a stationary sample.
5. Annual counts reflect filing volume, not only reliability change.

Citation

@misc{safiullah2026agenticfailures,
  title  = {Failures in Agentic AI: a five-dimension taxonomy and corpus of
             3,089 reported failures},
  author = {Safiullah, Khawaja},
  year   = {2026},
  note   = {Dataset and codebook, CC BY 4.0},
  url    = {https://huggingface.co/datasets/khawajasafiullah/failures-in-agentic-ai}
}

Maintenance

Regenerated from the project repository with python -m afs.cli hf. The card, the parquet schema and the counts above are produced by the same command, so a relabelled corpus cannot leave a stale number in the documentation.

Generated 2026-11-08T14:20:15Z.