Add Action Guardrails example (pre‑action/just‑in‑time authorization before tool calls)

[uchibeke]

I'd like to contribute an "Action Guardrails" example that complements existing input/output guardrails by enforcing authorization immediately before any effectful tool/API call (refunds, trades, data exports). This runs after the agent chooses an action but before side effects, enforcing agent identity (passport) and policy, failing closed by default, and emitting an auditable, immutable and verifiable decision ID.

This addresses a distinct concern from input/output guardrails:

• Input/output guardrails: Protect against malicious/unsafe data
• Action guardrails: Enforce business policies and identity on actions

The example demonstrates a generic, framework-agnostic pattern. It uses APort as the implementation example, but developers can adapt it to use their own authorization service or policy enforcement mechanism.

How APort Action Guardrails Fit Into the Flow

sequenceDiagram
    participant User
    participant InputGuardrail as Input Guardrails<br/>(Sanitize Prompt)
    participant Agent
    participant ActionGuardrail as APort Action Guardrails<br/>(Pre-Action Authorization)
    participant Tool as Effectful Tool/API<br/>(Refund, Trade, Export)
    participant OutputGuardrail as Output Guardrails<br/>(Validate Response)

    User->>InputGuardrail: Raw prompt
    InputGuardrail->>InputGuardrail: Sanitize & validate
    InputGuardrail->>Agent: Clean prompt
    Agent->>Agent: Process & decide action
    Note over Agent: Agent chooses to:<br/>execute refund/trade/export
    Agent->>ActionGuardrail: Verify(passport, policy, context)
    ActionGuardrail->>ActionGuardrail: Check agent identity<br/>Enforce policy limits<br/>Validate context
    alt Authorization: ALLOW
      ActionGuardrail-->>Agent: Decision(allow=true, decision_id)
      Agent->>Tool: Execute action
      Tool-->>Agent: Result
      Agent->>OutputGuardrail: Response
      OutputGuardrail->>OutputGuardrail: Validate & transform
      OutputGuardrail-->>User: Safe response
    else Authorization: DENY
      ActionGuardrail-->>Agent: Decision(allow=false, reasons)
      Agent->>OutputGuardrail: Error response
      OutputGuardrail-->>User: Safe decline + rationale
    end

Loading

Three Types of Guardrails Comparison

flowchart TB
    subgraph "Input Guardrails"
        A1[User Prompt] --> A2[Sanitize & Validate]
        A2 --> A3[Agent Receives<br/>Clean Prompt]
        style A1 fill:#0277bd,stroke:#01579b,stroke-width:2px,color:#fff
        style A2 fill:#0288d1,stroke:#01579b,stroke-width:2px,color:#fff
        style A3 fill:#039be5,stroke:#01579b,stroke-width:2px,color:#fff
    end

    subgraph "Action Guardrails (APort)"
        B1[Agent Decides Action] --> B2[APort Verify<br/>Pre-Action Auth]
        B2 --> B3{Allow?}
        B3 -->|Yes| B4[Execute Tool]
        B3 -->|No| B5[Deny + Reasons]
        style B1 fill:#e65100,stroke:#bf360c,stroke-width:2px,color:#fff
        style B2 fill:#ff6f00,stroke:#bf360c,stroke-width:2px,color:#fff
        style B4 fill:#ff8f00,stroke:#bf360c,stroke-width:2px,color:#fff
        style B5 fill:#e65100,stroke:#bf360c,stroke-width:2px,color:#fff
    end

    subgraph "Output Guardrails"
        C1[Tool Result] --> C2[Validate & Transform]
        C2 --> C3[Safe Response<br/>to User]
        style C1 fill:#6a1b9a,stroke:#4a148c,stroke-width:2px,color:#fff
        style C2 fill:#7b1fa2,stroke:#4a148c,stroke-width:2px,color:#fff
        style C3 fill:#8e24aa,stroke:#4a148c,stroke-width:2px,color:#fff
    end

    A3 --> B1
    B4 --> C1
    B5 --> C1

Loading

APort Authorization Flow Detail

flowchart LR
    subgraph "Agent Runtime"
        A[Agent] -->|1. Action Selected| B[Build Context]
        B -->|2. Context| C[APort Client]
    end

    subgraph "APort Service"
        C -->|3. Verify Request| D[Policy Evaluator]
        D --> E{Check Passport}
        E -->|Valid| F{Check Limits}
        E -->|Invalid| G[Deny]
        F -->|Within Limits| H{Check Policy Rules}
        F -->|Exceeds Limits| G
        H -->|Pass| I[Allow + Decision ID]
        H -->|Fail| G
    end

    subgraph "Tool Execution"
        I -->|4. Decision| J{Authorized?}
        J -->|Yes| K[Execute Tool]
        J -->|No| L[Return Error]
        K --> M[Tool Result]
        L --> M
    end

    style D fill:#2e7d32,stroke:#1b5e20,stroke-width:2px,color:#fff
    style I fill:#388e3c,stroke:#1b5e20,stroke-width:2px,color:#fff
    style G fill:#c62828,stroke:#b71c1c,stroke-width:2px,color:#fff
    style K fill:#1565c0,stroke:#0d47a1,stroke-width:2px,color:#fff

Loading

Proposed contribution:

• Add examples/agent_patterns/guardrails_action.py showing a small wrapper that verifies via an injected client before executing a tool; denies with reasons when allow=false; attaches decision metadata to logs/trace.
• Keep it SDK‑agnostic; show a minimal context shape (operation, resource, amount/currency, user/tenant, region).
• The example uses policies based on the Open Agent Passport (OAP) v1.0 specification for standardized policy evaluation.

References:

• Agents Python guardrail examples for analogy: input_guardrails.py, output_guardrails.py
  • https://github.com/openai/openai-agents-python/blob/main/examples/agent_patterns/input_guardrails.py
  • https://github.com/openai/openai-agents-python/blob/main/examples/agent_patterns/output_guardrails.py
• Prior art PR: https://github.com/openai/openai-guardrails-python/pull/34
• Microsoft Agent Framework discussion: https://github.com/microsoft/agent-framework/discussions/1701
• Standards reference: Open Agent Passport (OAP) v1.0 Specification — Runtime trust specification for AI agent authorization
• Policy examples: OAP Policy Packs — Standardized policy implementations for financial transactions, data access, and more

If maintainers agree, I'll open a tiny PR with one runnable example + short README.

[seratch]

Thanks for sharing your idea! It sounds great for developers. However, we’d like to keep this repo focused on our SDK support. The example you suggested could be better suited for your own repo. So, please start by sharing the idea there and feel free to add OpenAI Agents SDK examples. Once that’s ready, we’d really appreciate it if you could share the resources here as well.