asago (AI Safety And Governance Orchestration) is an open-source community that aims to automate the journey from AI governance policy to production-ready, safely deployed AI systems — bridging the gap between compliance teams, AI engineers, and infrastructure operators.
The asago project is in its formation stage, however you can check in on our development at the repos below. As we release more, they will be added below.
Find out more about asago at https://asago.ai
Asago turns AI governance policies into concrete tests, runs those tests against an agent, and recommends fixes.
The pipeline has two stages. The first stage reads your policies and produces test scenarios. The second stage evaluates an agent against those scenarios and recommends guardrails or configuration changes.
The Policy Mapper reads your policy documents and produces a set of risks. The Scenario Generator then turns each risk into a technology-agnostic test scenario, based on the agent capabilities.
flowchart LR
Policy{{"Policies"}} --> PolicyMapper["Policy Mapper"]
PolicyMapper --> Risks{{"Risks"}}
Risks --> ScenarioGen["Scenario\nGenerator"]
ScenarioGen --> Scenarios{{"Scenarios"}}
UseCaseContext{{"Use Case Context"}} -->|optional| PolicyMapper
AgentDesc{{"Agent Description"}} --> ScenarioGen
classDef data fill: #dbeafe,stroke: #3b82f6,color: #1e3a5f
classDef process fill: #fef9c3,stroke: #eab308,color: #713f12
class Policy,Risks,Scenarios,UseCaseContext,AgentDesc data
class PolicyMapper,ScenarioGen process
Blue nodes represent data. Yellow nodes represent functional components.
The Artifact Generator turns the scenarios into framework-specific test artifacts, through the Midojo and Garak adapters. EvalHub runs these tests against the agent endpoint and produces metrics. The Recommender then reviews the results and produces guardrail or configuration recommendations.
The user applies the recommendations and re-runs the evaluation against the same scenarios. This loop continues until all risks pass.
flowchart LR
AgentEndpoint{{"Agent Endpoint"}}
AgentEndpoint --> ArtifactGen["Artifact Generator"]
ArtifactGen --> RunArtifact{{"Run Artifact"}}
RunArtifact --> EvalHub["EvalHub"]
EvalHub --> Metrics{{"Metrics"}}
Scenarios{{"Scenarios"}} --> ArtifactGen
Scenarios --> RecommenderGroup
Metrics --> RecommenderGroup
GuardrailCatalog{{"Guardrail\nCatalog"}} --> GuardrailsRec
subgraph RecommenderGroup ["Recommender"]
GuardrailsRec["Guardrails\nRecommender"]
ConfigRecommender["Configuration\nRecommender"]
end
GuardrailsRec & ConfigRecommender --> Recommendations{{"Recommendations"}}
classDef data fill: #dbeafe,stroke: #3b82f6,color: #1e3a5f
classDef process fill: #fef9c3,stroke: #eab308,color: #713f12
class Scenarios,Metrics,Recommendations,GuardrailCatalog,AgentEndpoint,RunArtifact data
class ArtifactGen,EvalHub,GuardrailsRec,ConfigRecommender process
Blue nodes represent data. Yellow nodes represent functional components.
| Repo | Description |
|---|---|
| Policy Mapper | Reads your AI policy documents and produces a structured list of identified AI risks, with evidence and cross-taxonomy mappings. Use by future downstream asago components |
| Scenario Generator | Translates the identified risks into technology-agnostic test scenarios, based on the agent capabilities. Consumed by downstream asago components. |
| Artifact Generator | Translates the test scenarios into framework-specific test artifacts through the Midojo and Garak adapters. |
| Repo | Description |
|---|---|
| Garak (fork) | Midstream fork of Nvidia Garak for automated red-teaming of LLMs. Used as an adapter by the Artifact Generator. |
| Midojo | Red-teams AI agents by hiding malicious payloads in the data they fetch, testing whether the agent takes harmful actions as a side effect of doing legitimate work. Will be orchestrated by future asago components. |
| Repo | Description |
|---|---|
| Examples | Jupyter notebooks to get you started with asago components. |