Moving from Capabilities to Responsibilities in High-Stakes Agentic AI

From capabilities to responsibilities

Artur Huk introduces the Responsibility-Oriented Agent (ROA) architecture. This system replaces probabilistic prompt-based guardrails with a deterministic execution kernel to manage high-stakes side effects.

Why This Matters

Current enterprise AI relies on ‘capabilities’ (what an agent can do) and Human-in-the-Loop (HITL) reviews, which creates a scalability trap where alert fatigue leads humans to approve payloads without review. In production systems moving money or changing infrastructure, relying on LLM alignment and prompts is insufficient; a hard boundary between an agent’s epistemic intent and the system’s executive action is required to prevent catastrophic failures.

Key Insights

• The Actor model (introduced by Carl Hewitt in 1973) provides the foundation for ROAs by treating agents as bounded entities that communicate via messages rather than open capability sets.
• Responsibility Contracts replace natural language prompts with machine-readable code (e.g., max_tiv: 3000000) to ensure boundaries are enforced deterministically at runtime.
• Epistemic Isolation separates ‘Explain’ narratives from ‘Policy’ proposals, ensuring that free text is used for auditing while only structured, typed claims trigger execution.
• Decision Flow IDs (dfid) create an immutable reconstruction primitive, linking the T0 context snapshot to the validation outcome and final execution receipt.

Working Examples

Example of a versioned Responsibility Contract defining hard authority boundaries for an underwriting agent.

agents:
- agent_id: "underwriter_agent"
version: "1.0.0"
created_by: "[email protected]"
created_at: "2025-02-17T10:00:00Z"
mission: |
You are an insurance underwriter. Analyze the client application and propose
a policy. Base premium on Total Insured Value (TiV) at ~2% of TiV, capped at max_tiv.
NEVER propose for Fireworks or CryptoMining industries - these are prohibited.
contract:
role: EXECUTOR
max_tiv: 3000000
prohibited_industries: ["Fireworks", "CryptoMining"]
escalate_on_uncertainty: 0.65

Practical Applications

• Use case: London commercial market underwriting where agents propose quotes up to a £10M contract cap and escalate only when values exceed that limit. Pitfall: Using prompts for position sizing, which can be bypassed via prompt injection or model drift.
• Use case: Logistics rerouting focusing on minimizing SLA penalties using low-cost carriers within a fixed budget. Pitfall: Stateless inference calls leading to infinite rejection loops because the agent lacks memory of prior failed proposals.

References:

• https://stackoverflow.blog/2026/06/19/dispatches-from-o-reilly-from-capabilities-to-responsibilities/