id string | inputs dict | tags list | expected_tools list | toolset_ref string | expected_no_tools bool |
|---|---|---|---|---|---|
ex_security_01 | {
"query": "User account_42 had 5 failed login attempts in the last hour"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_02 | {
"query": "Someone is trying to access account_77 from a new IP in another country"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_03 | {
"query": "Customer reports they cannot log in and admin panel was probed 8 times"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_04 | {
"query": "Our SOC alerted on user_99 - possible account takeover"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_05 | {
"query": "alice@example.com tried to access the admin dashboard 8 times in 5 minutes"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_06 | {
"query": "Suspicious activity on customer_55 - multiple sessions from different countries"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_07 | {
"query": "Alert: customer_18 reported their account was compromised; lock the session"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_08 | {
"query": "Possible API key leak detected for tenant_3 - multiple unusual calls"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_09 | {
"query": "Customer support flagged that user_404 wallet was drained in 1 minute"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_10 | {
"query": "Mass-credential-stuffing attempt against /login from 200 IPs"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_11 | {
"query": "Internal scan found customer_7 PII exposed in logs"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_security_12 | {
"query": "Confirmed phishing email targeting customer_92 - they clicked a link"
} | [
"security"
] | [
{
"tool_name": "notify_security_team",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_01 | {
"query": "Where is order #12345 for customer_42?"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_02 | {
"query": "Customer customer_77 is asking about their refund - please look up their orders"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_03 | {
"query": "Send a follow-up email to billing@acme.com about ticket 203"
} | [
"routine"
] | [
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_04 | {
"query": "Please look up customer_88 last 3 orders"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_05 | {
"query": "Confirmation email needed for order #98765 to ops@acme.com"
} | [
"routine"
] | [
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_06 | {
"query": "Customer customer_31 wants their order history emailed to billing@acme.com"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_07 | {
"query": "What was the most recent purchase from customer_12?"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_08 | {
"query": "Send a welcome email to new customer onboarding@example.com"
} | [
"routine"
] | [
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_09 | {
"query": "Look up orders for customer_50 then email a summary to support@acme.com"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_10 | {
"query": "Customer customer_5 asked about the status of order #66677"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_11 | {
"query": "Pull up the order list for customer_63 - they want a refund estimate"
} | [
"routine"
] | [
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_routine_12 | {
"query": "Send a return-shipping label to customer_19 at returns@acme.com for order #54321"
} | [
"routine"
] | [
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_refund_01 | {
"query": "Issue a $49.99 refund on order #ord_12345 for damaged item"
} | [
"refund"
] | [
{
"tool_name": "issue_refund",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_refund_02 | {
"query": "Customer customer_42 wants $120 back on order #ord_99887, says product never arrived. Email confirmation to customer_42@example.com."
} | [
"refund"
] | [
{
"tool_name": "issue_refund",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_refund_03 | {
"query": "Look up customer_77 by email and refund their last order $200 for late delivery"
} | [
"refund"
] | [
{
"tool_name": "lookup_customer",
"match_strategy": "subset"
},
{
"tool_name": "issue_refund",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_refund_04 | {
"query": "Refund order #ord_55543 ($30) for customer_8 and mark it cancelled"
} | [
"refund"
] | [
{
"tool_name": "issue_refund",
"match_strategy": "subset"
},
{
"tool_name": "update_order_status",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_refund_05 | {
"query": "Process a $75 refund on order #ord_44231 - reason: duplicate charge"
} | [
"refund"
] | [
{
"tool_name": "issue_refund",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_customer_lookup_01 | {
"query": "Look up the customer with email alice@example.com"
} | [
"customer_lookup"
] | [
{
"tool_name": "lookup_customer",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_customer_lookup_02 | {
"query": "Find the customer record for customer_99"
} | [
"customer_lookup"
] | [
{
"tool_name": "lookup_customer",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_customer_lookup_03 | {
"query": "Pull up bob@example.com profile and email them a security tip at the same address"
} | [
"customer_lookup"
] | [
{
"tool_name": "lookup_customer",
"match_strategy": "subset"
},
{
"tool_name": "send_email",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_customer_lookup_04 | {
"query": "Who is customer_31? Show me their orders too."
} | [
"customer_lookup"
] | [
{
"tool_name": "lookup_customer",
"match_strategy": "subset"
},
{
"tool_name": "search_orders",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_customer_lookup_05 | {
"query": "Find customer info for support@example.com"
} | [
"customer_lookup"
] | [
{
"tool_name": "lookup_customer",
"match_strategy": "subset"
}
] | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | null |
ex_text_01 | {
"query": "What is your refund policy?"
} | [
"text_only"
] | null | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | true |
ex_text_02 | {
"query": "How do I reach billing?"
} | [
"text_only"
] | null | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | true |
ex_text_03 | {
"query": "What hours is your support team available?"
} | [
"text_only"
] | null | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | true |
ex_text_04 | {
"query": "Do you ship internationally?"
} | [
"text_only"
] | null | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | true |
ex_text_05 | {
"query": "What payment methods are accepted?"
} | [
"text_only"
] | null | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | true |
ex_text_06 | {
"query": "Where can I find your terms of service?"
} | [
"text_only"
] | null | sha256:ad3f07832877239336d594c7c1626197b39b9eb4c178842304c8754e94323d67 | true |
Agent Model-Migration Golden Suite (tool routing)
A small, ready-to-run golden suite for one question: if I switch the model behind my agent, does it still call the right tools?
40 inbound customer-support messages, each labelled with the tools a correct agent must call (or must not call). Run it against your current model and a candidate model and compare tool routing case by case.
It is the agent example from the open-source EvalShift CLI (Apache-2.0), packaged as a standalone dataset.
Why tool routing?
When teams swap models (GPT → Claude, Gemini 2.5 → 3.x, a closed API → an open-weight Qwen or Llama), aggregate benchmark scores often barely move. What breaks is behaviour the app depends on:
- the new model skips a required tool (e.g. never pages security on an account-takeover message),
- it calls tools in the wrong order (sends the confirmation email before issuing the refund),
- or it calls tools when it should just answer ("what is your refund policy?").
Each row here targets one of those failure modes.
Structure
Slice (tags) |
Rows | What it checks |
|---|---|---|
security |
12 | notify_security_team must be called for failed logins, takeover attempts, key leaks, etc. |
routine |
12 | Order lookups and status updates call only the tools needed |
refund |
5 | issue_refund (plus lookup_customer / send_email / update_order_status where asked) |
customer_lookup |
5 | lookup_customer when only an email or id is given |
text_only |
6 | No tool calls: the model should answer directly |
Example row:
{"id": "ex_security_01",
"inputs": {"query": "User account_42 had 5 failed login attempts in the last hour"},
"tags": ["security"],
"expected_tools": [{"tool_name": "notify_security_team", "match_strategy": "subset"}],
"toolset_ref": "sha256:ad3f0783..."}
Fields:
inputs.query: the inbound message, substituted into the system prompt inprompts.pytags: slice labelexpected_tools: tools that must appear in the model's calls (subsetmatch)expected_no_tools:truewhen the correct behaviour is to answer without calling any tooltoolset_ref: points to the tool definitions intoolsets/(same six tools astools.yaml)
Files:
| File | Purpose |
|---|---|
golden.jsonl |
The 40 cases |
prompts.py |
Agent system prompt with routing rules |
tools.yaml / toolsets/*.json |
Six tool schemas: search_orders, lookup_customer, issue_refund, update_order_status, send_email, notify_security_team |
evalshift.yaml |
Ready-made EvalShift config (tool_selection evaluator, slices from tags) |
Load it
from datasets import load_dataset
ds = load_dataset("Liveta/agent-migration-golden-suite", split="test")
print(ds[0]["inputs"]["query"], ds[0]["expected_tools"])
The format is plain JSONL, so you can score it with your own harness: send prompts.py plus the tools to each model and check the called tool names against expected_tools / expected_no_tools.
Run a model-vs-model comparison with EvalShift
pip install evalshift # Python 3.11+
huggingface-cli download Liveta/agent-migration-golden-suite --repo-type dataset --local-dir agent-suite
cd agent-suite
export GOOGLE_API_KEY=...
evalshift run --yes --from gemini-2.5-flash --to gemini-3.1-flash-lite-preview
RUN_ID=$(ls -t .evalshift/runs/ | head -1)
evalshift evaluate "$RUN_ID"
evalshift analyze "$RUN_ID"
evalshift report "$RUN_ID" --open
Model calls go through LiteLLM, so any LiteLLM model id works as either side, including self-hosted open-weight models (hosted_vllm/..., ollama/...). The bundled evalshift.yaml uses Gemini models for the defaults and the judge; edit it to match the providers you have keys for.
The report shows per-slice tool-routing regressions, so a candidate that is fine on routine traffic but stops paging security on takeover attempts can't hide behind a good average.
To build a suite like this from your own agent's real traffic instead of hand-written cases, use the capture SDK (evalshift-sdk, MIT).
Limitations
- Small (40 rows) and synthetic: it's a smoke test for tool-routing regressions, not a benchmark or leaderboard.
- Single domain (customer-support routing) with six tools.
- Checks which tools are called, not argument values or final answer quality.
Related work
- MUSCLE: A Model Update Strategy for Compatible LLM Evolution (negative flips across model updates): https://arxiv.org/abs/2407.09435
- ReCatcher: Towards LLMs Regression Testing for Code Generation: https://arxiv.org/abs/2507.19390
- PromptBridge: Cross-Model Prompt Transfer for Large Language Models ("model drifting" when switching models): https://arxiv.org/abs/2512.01420
- Berkeley Function Calling Leaderboard: https://huggingface.co/datasets/gorilla-llm/Berkeley-Function-Calling-Leaderboard
Source and license
Taken from examples/agent/ in evalshift/evalshift-cli, Apache-2.0. More on model-migration testing: https://evalshift.dev
- Downloads last month
- 57