Datasets:
metadata
license: mit
tags:
- agentic-ai
- error-recovery
- tool-calling
- synthetic-data
- evaluations
- function-calling
task_categories:
- text-generation
size_categories:
- n<1K
configs:
- config_name: default
data_files: agentic_workflows_sample.jsonl
Agent-Resilience-Eval: Production API Failure & Auto-Recovery Benchmark
An enterprise evaluation benchmark designed to measure and fine-tune Agentic AI fault tolerance.
Most production agents crash, loop, or hallucinate when third-party APIs return non-standard errors (HTTP 409 stale locks, HTTP 412 ETag conflicts, HTTP 429 rate limit backoffs, and OAuth token invalidations). This benchmark provides 64 multi-turn recovery trajectories across mission-critical systems.
Trajectory Schema
Every record captures a full fault-and-remediation cycle:
{
"task_id": "DEVO-D1FCDF",
"domain": "DevOps / GitHub API",
"user_intent": "Attempt to squash and merge a pull request on GitHub that has encountered a conflict.",
"initial_tool_call": {
"tool": "github_api_call",
"parameters": {"endpoint": "pulls/114/merge", "method": "POST"}
},
"error_encountered": {
"status_code": 409,
"error_type": "Conflict",
"message": "The pull request's head branch is out of date."
},
"recovery_reasoning": "The branch has diverged. The agent synchronizes the head branch via rebase before retrying the squash merge.",
"corrective_tool_call": {
"tool": "github_api_call",
"parameters": {"endpoint": "pulls/114/rebase", "method": "POST"}
},
"subsequent_tool_call": {
"tool": "github_api_call",
"parameters": {"endpoint": "pulls/114/merge", "method": "POST"}
},
"final_state": "SUCCESS: The pull request has been successfully merged after resolving the merge conflict.",
"metadata": {"difficulty": "medium", "error_class": "Conflict", "protocol": "REST"}
}
Domains Covered (15+ Production APIs)
FinTech & Payments: Stripe (3DS challenge handling), Plaid (MFA update tokens), Square, PayPal, Coinbase.
DevOps & Cloud: GitHub Actions, Kubernetes RBAC, Terraform Cloud lock releases, AWS S3 / KMS, Docker Registry, Vercel deployments.
Databases & Cache: Supabase RLS policies, PostgreSQL pool exhaustion, Redis memory eviction, Snowflake SQL timeouts, Pinecone vector limits.
SaaS & CRM: Salesforce lock contention, HubSpot search pagination limits, Zendesk triage routing, Slack rate-limit backoffs.
Health & Auth: HL7 FHIR clinical record ETag conflicts, Okta burst rate limits, Auth0 token expirations.
Commercial Licensing & Custom Batches
Free Tier (This Repo): 64 verified sample trajectories under the MIT License.
Enterprise Benchmark Tier ($299): 500 production edge-case trajectories pre-packaged with Pytest evaluation harnesses.
Custom Schema Generation ($699): 500 bespoke test cases matched specifically to your internal OpenAPI/Swagger tool definitions.
To order custom eval batches:
Contact: hello@truzot.com