01 · find failures
Maul
How does the agent behave under failure?
An adversarial OpenAI-compatible proxy. Point the agent's
base_url at Maul to inject controlled failures and record
how the workflow responds.
Artifact reliability_report.json
Explore Maul on GitHub ↗
terminal · maul
$ maul test --config maul.yaml \
--agent "python agent.py"
scenario force_429
faults_injected 3
recovered 2
report written artifacts/reliability_report.json
02 · measure quality
Holds
Did the agent actually solve the job?
A suite-first evaluation harness. Run representative tasks repeatedly,
grade real outcomes, and stop regressions before they reach production.
Artifacts holds.yaml · baseline.json
Explore Holds on GitHub ↗
terminal · holds
$ holds run --suite holds.yaml
tasks 12
pass_rate 96%
consistency 92%
threshold passed
$ holds compare --baseline baseline.json
03 · prevent incidents
Vigil
Are unsafe or expensive behaviors stopped?
A runtime policy SDK for explicit operational controls. Define budgets,
allowlists, circuit breakers, routing, and redacted audit behavior.
Artifacts vigil.yaml · audit.jsonl
Explore Vigil on GitHub ↗
vigil.yaml
budgets:
max_llm_calls_per_workflow: 12
max_cost_usd_per_workflow: 0.75
controls:
- type: circuit_breaker
condition: repeated_equivalent_request
max_occurrences: 3