CortexOps

Reliability infrastructure for AI agents.
Evaluate · Observe · Operate — for LangGraph, CrewAI, and AutoGen.

What's New in v0.5.1

Prompt version tracking

tracer = CortexTracer(project="payments-agent")
tracer.set_prompt("refund-system", "v3")   # every trace is tagged refund-system@v3
graph = tracer.wrap(your_langgraph_app)

with tracer.use_prompt("refund-system", "v4-candidate"):
    graph.invoke({"messages": [...]})       # tagged v4-candidate, then back to v3

Filter hosted traces with GET /v1/traces?project=...&prompt_name=refund-system&prompt_version=v3.

pytest eval gates

# test_agent_quality.py — pip install "cortexops[pytest]"
import cortexops

@cortexops.eval_gate("golden_v1.yaml", fail_on="task_completion < 0.90")
def test_payments_agent():
    return build_agent()   # return the agent; the gate runs every case

The plugin loads automatically, prints an eval summary at the end of the pytest run, and marks gates with cortexops_eval so pytest -m "not cortexops_eval" skips them.

Sampling configuration

export CORTEXOPS_SAMPLE_RATE=0.1   # trace 10% of successful runs; failures are always captured

Or pass CortexTracer(sample_rate=0.1, sample_errors=True).

What's New in v0.4.0

LLM-as-judge evaluation

from cortexops.judge import LLMJudge

judge = LLMJudge(api_key="sk-...")
result = judge.evaluate(
    case_id="case-001",
    input="Process refund for order #4821",
    output="Refund of $49.99 approved and processed.",
    rubric="task_completion",
)
print(result.score, result.passed, result.reasoning)

Golden dataset API

from cortexops.dataset import GoldenDataset

ds = GoldenDataset(name="refund-agent-v1")
ds.add(input="Refund order #4821", expected="refund_approved")
ds.add(input="Cancel subscription", expected="subscription_cancelled")
ds.save("datasets/refund_agent.yaml")

results = ds.run(agent=your_agent, fail_on="task_completion < 0.90")

CI/CD eval gate

cortexops eval run \
  --dataset datasets/refund_agent.yaml \
  --judge \
  --fail-on "task_completion < 0.90"
# Exit code 1 if regression detected — drop into GitHub Actions

The problem

You deployed an agent. You have no idea if it regressed overnight.

No standard eval format. No failure traces. No CI gate before the next prompt change ships.
CortexOps fixes that.

Quickstart

pip install cortexops  # v0.4.0

from cortexops import CortexTracer, EvalSuite

# Wrap your LangGraph app — zero refactor required
tracer = CortexTracer(project="payments-agent")
graph  = tracer.wrap(your_langgraph_app)

# Run evaluations against a golden dataset
results = EvalSuite.run(
    dataset="golden_v1.yaml",
    agent=graph,
)

print(results.summary())
# CortexOps eval — payments-agent
#   Cases           : 9  (7 passed, 2 failed)
#   Task completion : 91.4%
#   Tool accuracy   : 97.0/100
#   Latency p50/p95 : 42ms / 187ms
#   Failed cases:
#     - escalation_router: tool_call_mismatch (score 41)

Golden dataset format

Define test cases in YAML. Run them locally or in CI.

# golden_v1.yaml
version: 1
project: payments-agent

cases:
  - id: refund_lookup_01
    input: "What is the status of refund REF-8821?"
    expected_tool_calls: [lookup_refund]
    expected_output_contains: ["approved", "REF-8821"]
    max_latency_ms: 3000

  - id: dispute_escalation_01
    input: "I was charged twice — this is unauthorized"
    expected_tool_calls: [classify_dispute, route_escalation]
    expected_output_contains: ["escalated"]
    max_latency_ms: 5000

CI eval gate

Add to .github/workflows/eval.yml:

- name: CortexOps eval gate
  run: |
    python examples/langgraph_payments/run_eval.py \
      --dataset golden_v1.yaml \
      --fail-on "task_completion < 0.90"

If the eval drops below threshold, the job exits non-zero and the PR is blocked.

Prefer pytest? Use @cortexops.eval_gate(...) and run the gate with the rest of your test suite.

Repo structure

cortexops/
├── sdk/                        # pip install cortexops  # v0.4.0
│   ├── cortexops/
│   │   ├── tracer.py           # CortexTracer — wraps LangGraph / CrewAI
│   │   ├── eval.py             # EvalSuite — golden dataset runner
│   │   ├── metrics.py          # task_completion, tool_accuracy, latency, hallucination
│   │   ├── models.py           # Pydantic data models
│   │   └── client.py           # HTTP client for hosted API
│   └── tests/
├── backend/                    # FastAPI + Celery + SQLite/Postgres
│   ├── app/
│   │   ├── main.py
│   │   ├── routers/            # /v1/evals, /v1/traces
│   │   ├── models/             # DB records + API schemas
│   │   └── worker/             # Celery async eval tasks
│   └── Dockerfile
├── frontend/                   # React + TypeScript dashboard
├── examples/
│   └── langgraph_payments/     # Full runnable demo
│       ├── agent.py
│       ├── golden_v1.yaml
│       └── run_eval.py
└── docker-compose.yml

Run the full stack locally

git clone https://github.com/ashishodu2023/cortexops
cd cortexops

# Start API + worker + Redis
docker compose up --build

# In another terminal — run the demo eval
cd examples/langgraph_payments
pip install -e ../../sdk/
python run_eval.py

# API docs at http://localhost:8000/docs
# Dashboard at http://localhost:3000

Supported frameworks

Built-in metrics

Add custom metrics by subclassing cortexops.Metric.

Contributing

git clone https://github.com/ashishodu2023/cortexops
cd cortexops/sdk
pip install -e ".[dev]"
pytest tests/ -v

See CONTRIBUTING.md. Issues labeled good first issue are a great place to start.

Citation

@software{cortexops2025,
  author  = {Ashish, et al.},
  title   = {CortexOps: Reliability Infrastructure for AI Agents},
  year    = {2025},
  url     = {https://github.com/ashishodu2023/cortexops},
}

License

MIT — see LICENSE.

cortexops.ai · Issues · Discussions