agentic-programming.md

Agentic Programming

Spanda supports safety-gated agentic programming with tool permissions, memory scopes, and approval gates.

Example

agent Planner {
    goal "Navigate safely";
    tools [camera, lidar, map];
    memory short_term;
    policy safe_only;

    can [ read(lidar), propose_motion ];

    plan {
        let proposal = planner.reason(goal);
        let action = safety.validate(proposal);
        return action;
    }
}

Rules

• Agents cannot directly execute actuators unless permitted and safety-gated
• High-risk actions require approval or SafeAction
• Empty can [] default-denies propose_motion and execute at runtime (Phase 32)
• Capability grant/deny events are written to the audit trail when configured (Phase 31)
• Reasoning traces captured for audit when audit is configured
• ActionProposal must pass through safety.validate before actuator execution
• Agent plan blocks with motion grants must return SafeAction — see typed-handler-io.md

Examples:

• examples/features/agent_capabilities.sd — populated can[]
• examples/features/agent_can_deny.sd — empty can[] denial

Runtime

spanda-ai provides AgentRuntime with mock and live AI paths. Capability enforcement runs in the interpreter.

See Architecture — AI Safety and Feature Status.