AgentDisruptBench

An evaluation methodology for measuring AI agent resilience under runtime tool-call disruptions.

Overview

AgentDisruptBench provides a systematic methodology, backed by 100 base tasks and variants across 4 domains, a systematic 20-type disruption taxonomy, and 9 disruption severity profiles, to study and measure how well LLM-based agents handle real-world tool failures.

Task Statistics

Domain	Standard	Adversarial	Impossible	Handover	Total
Retail	20	2	2	1	25
Travel	20	2	2	1	25
Finance	20	2	2	1	25
DevOps	20	2	2	1	25
Total	80	8	8	4	100

Disruption Taxonomy (20 Types)

Category	Types
Timing	timeout, latency
HTTP Status	http_429, http_401, http_403, http_500, http_502, http_503
Response Content	malformed_json, truncated, null_response, missing_fields, type_mismatch, schema_drift, wrong_data
Behavioral	intermittent, flapping, quota_exhausted, auth_expiry, cascading

Key Metric: R(k, ε, λ) Reliability Surface

• k-consistency: Pass rate across repeated seeds (same task, same profile)
• ε-robustness: Pass rate across task-wording variants
• λ-fault-tolerance: Pass rate across disruption profiles

Files

• tasks/retail.yaml — 20 retail domain tasks
• tasks/travel.yaml — 20 travel domain tasks
• tasks/finance.yaml — 20 finance domain tasks
• tasks/devops.yaml — 20 DevOps domain tasks
• tasks/adversarial.yaml — 8 adversarial trap tasks
• tasks/impossible.yaml — 8 impossible tasks
• tasks/handover.yaml — 4 handover tasks
• tasks/variants.yaml — 18 ε-robustness task variants
• profiles/ — 9 disruption profile definitions

Usage

pip install agentdisruptbench

from agentdisruptbench import TaskRegistry, DisruptionEngine
registry = TaskRegistry.from_builtin()
tasks = registry.filter(domain="retail", max_difficulty=3)

Citation

@inproceedings{agentdisruptbench2026,
  title={AgentDisruptBench: An Evaluation Methodology for AI Agent Resilience Under Runtime Tool-Call Disruptions},
  author={AgentDisruptBench Contributors},
  year={2026}
}

License

MIT