agentstress

agentstress

A stress-test benchmark for LLM agent frameworks. 100 scenarios engineered to provoke five failure modes from the MAST taxonomy, with graders that say which failure happened and whether the task was actually done correctly — two separate questions that often disagree.

It exists because of what it found: most of the difference between agent frameworks is the prompt scaffolding they wrap around your task, not their orchestration.

A single sentence CrewAI appends to every task — "you MUST return the actual complete content as the final answer, not a summary" — roughly doubles how often an agent acts on an underspecified request instead of asking, on both a 7B model and gpt-5.4-mini. Full results: docs below.

Install

pip install agentstress                 # scenarios, graders, CLI
pip install "agentstress[langgraph]"    # + the framework you want to test
pip install "agentstress[all]"          # + crewai, openai-agents, judge

Runs against a local model through Ollama by default, so you can use the whole benchmark without an API key.

Start here: what is your framework actually sending?

agentstress capture --framework crewai --scenario FAQ-12

prints the exact first request, separating your task text from everything the framework added. Most developers have never seen it.

Run the benchmark

ollama pull qwen2.5:7b-instruct

agentstress list --mode FAQ                       # see what is provoked
agentstress run  --framework langgraph --mode FAQ --trials 2 --out runs/
agentstress grade --runs runs/                    # step repetition: free, deterministic
agentstress grade --runs runs/ --judge            # all modes: billed LLM judge
agentstress report --runs runs/

Use a hosted model with --model openai:gpt-5.4-mini (needs OPENAI_API_KEY). The judge needs ANTHROPIC_API_KEY; it is never from the same model family as the agent under test.

What the suite measures

Design rules that make the numbers mean something:

• Provocation, not instruction. No scenario mentions failure modes, grading, or prohibitions. A test enforces this across all 100 scenarios.
• The artifact is the action. When the chat reply and the ticket disagree, the ticket is graded. Small models often say the right thing and write [placeholder] into the artifact.
• Controls. 8 scenarios where the failure is not available. A grader that flags them is over-flagging.
• Two axes. Failure mode and task correctness are scored separately, and they disagree often enough to matter.
• Scenario, not trial, is the unit. Trials of a scenario agree 81% of the time; treating them as independent inflates significance roughly fourfold.

Grading

Step repetition is graded deterministically from the trace (three tiers, with exemptions for legitimate retries). The other four modes use an LLM judge (claude-opus-5), validated three times against blind human labels: 96% and 96.7% (κ=0.93) on qwen traces, 87.5% (κ=0.75) on GPT traces, where a known leniency is documented and bounded.

Results and data

Older documents in the repository root record how the study developed, including a discarded run whose prompts leaked grader rationale (PHASE_1_REWRITE_LOG.md).

Reproduce

pip install -e ".[all,dev]"
python -m tests.test_prompt_hygiene   # scenario prompts stay agent-facing
python -m tests.test_grader           # deterministic grader
python -m tests.test_correctness      # ground-truth checks
python -m tests.test_world            # mock world is pinned
python analyze_100.py                 # main statistics, no API calls
python paper/make_figures.py          # figures + numbers for the paper

Grading everything again costs money; every graded verdict is cached in results/*/judge_cache.json, so the statistics above re-run for free.

Limitations

• Measured with LangGraph 1.2.11, CrewAI 1.15.17, openai-agents 0.22.2 in September 2026. Framework defaults change; re-run capture on your versions.
• Two models (qwen2.5 7B, gpt-5.4-mini). On the stronger model the repetition and termination effects mostly disappear; the clarification effect does not.
• Part of CrewAI's scaffolding is wording this harness supplied for fields CrewAI requires; the ablations separate the two, and both carry the effect.
• One synthetic mock environment with nine tools.
• Only 15 of 100 scenarios involve an agent-to-agent handoff, and they show no framework difference. MAST's inter-agent failure modes are not covered.

Citation

@misc{agentstress2026,
  title  = {The Prompt Is the Framework: Agent Framework Scaffolding,
            Not Orchestration, Drives Failure Rates},
  author = {Patel, Arth},
  year   = {2026},
  note   = {Preprint}
}

License

MIT — see LICENSE.