AgentOrchBench

A benchmark of silent API failures for LLM agents.

An API can fail silently. It ignores part of a request, such as a misnamed filter, and still returns HTTP 200 with a plausible but wrong result. Nothing in the response shows the fault, so agents report the wrong number. AgentOrchBench measures how often agents fall for this and whether a guard can stop them.

Every task has an exact answer. Each count task also lists what the API returns under each silent fault, so you can see which faults change the answer and which ones an exact check can detect.

• Code, services and the SilentGuard harness: https://github.com/Sajib-006/silentguard
• Paper: Invariant-Aware Harness Enforcement to Protect LLM Agents from Silent API Failures

Configs

Config	Rows	Contents
count (default)	36	Count questions with 2 to 3 conditions, the main suite of the paper
lists	30	Questions whose answer is the list of matching record IDs
retail	30	Count questions on a second domain (retail orders)
workflows	100	Multi-step workflows across eight services
faults	8	The silent fault types, the API bug each is modeled on, and whether the exact count check sees it
catalog	400	The synthetic catalog records the count and list questions are asked about

from datasets import load_dataset

tasks = load_dataset("Sajib-006/AgentOrchBench", "count", split="test")
print(tasks[0]["prompt"], tasks[0]["answer"])

Fields (count and retail)

Field	Meaning
task_id	Task identifier
prompt	The question given to the agent. The agent must finish with count=<n>
answer	The exact correct count
conditions	The equality conditions in the question (field, value)
num_conditions	2 or 3
fault_effects	For each fault: the count the faulty API returns, whether it changes the answer, and whether the exact count check fires

Silent faults

Fault	API behavior	Modeled on	Exact count check
drop_last	drops the last filter	a misnamed filter	fires
cap_total	reports at most 5 matches	page size returned as the total	fires
or_semantics	joins conditions with OR	a faulty query builder	fires
drop_field	ignores any filter on one field	an unindexed field	fires
ignore_negation	drops NOT from a condition	no NOT support	false alarm
off_by_one	adds one to the count	a range bug	fires
prefix	matches by prefix	loose matching	cannot see
consistent_lie	lies the same way on every query	an adversarial API	cannot see

The first six are the fault types of the main results. The last two keep the count identity true by design and test the limits of exact checks.

How to run agents on it

The questions are answered through local services that you start from the code repository. A faulted service still answers HTTP 200, so the agent sees nothing unusual:

from agentorchbench import CatalogClient, load_tasks, score, start_services

with start_services("bio") as svc:
    api = CatalogClient(svc)
    api.set_fault("drop_last")
    for task in load_tasks("count"):
        answer = my_agent(task.prompt, tool=api.count)
        print(task.id, score(task, answer))   # correct, wrong, unknown or no_answer

Report wrong answers under faults, correct answers on fault-free tasks, and abstentions ("cannot be determined") separately. An agent that always abstains is never wrong but never useful.

Main results

Across 7 open models, unguarded agents report a wrong count in 74.0% of faulted runs. Checklists, reviewer models and LLM judges leave over 70% wrong. SilentGuard, which tests each result against exact invariants and withholds values that break them, leaves 1.7% wrong on the fault types its checks target, with no loss on fault-free tasks. See the paper for all methods, models and limits.

Data

All records are synthetic. Values are built from public dataset names (PBMC 3k, Allen Brain Atlas, GEO) and generic labels. No personal, patient or company data is included.

Citation

@article{dip2026silentguard,
  title  = {Invariant-Aware Harness Enforcement to Protect LLM Agents from Silent API Failures},
  author = {Dip, Sajib Acharjee and Ouyang, Brian and Motlhabi, Lala},
  year   = {2026}
}