Question 1

What is an AI agent?

Accepted Answer

An AI agent is a software system that combines a foundation model (typically an LLM) with tool access, memory, and a control loop to autonomously interpret instructions, make decisions, and take actions. Unlike simple chatbots, agents can execute multi-step tasks, call external APIs, and modify external state.

Question 2

What is the Agentic Architecture Framework?

Accepted Answer

The Agentic Architecture Framework (AAF) is a community-driven, vendor-agnostic set of architecture guidance for building safe, reliable, and governed agentic AI systems. It defines six pillars (Security, Reliability, Cost, Operations, Performance, Sustainability) and two cross-cutting foundations (Context Optimization, Autonomy Governance).

Question 3

What is an agentic control loop?

Accepted Answer

The agentic control loop is the repeating cycle an agent follows: Trigger → Interpret Context → Decide → Act → Observe Results → Verify → [Adapt / Stop]. A simplified form is Trigger → Decide → Act → Verify. Every architectural decision in the framework maps to a stage of this loop.

Question 4

How do you secure AI agents?

Accepted Answer

Securing AI agents requires a systems-security approach: least-privilege tool access, tool gateways that validate every call, prompt injection defences, privilege separation between reasoning and execution, and applying the CIA triad (Confidentiality, Integrity, Availability) to every component of the agent system.

Question 5

What is prompt injection and how do you defend against it?

Accepted Answer

Prompt injection is an attack where malicious instructions are hidden in data the agent processes, causing it to deviate from its intended behaviour. Defences include instruction hierarchy enforcement, input sanitisation, output filtering, separating trusted instructions from untrusted data, and never allowing user-supplied content to modify the system prompt.

Question 6

How do you manage AI agent costs?

Accepted Answer

Agent costs are controlled through token budgets per task, compute quotas, model routing (using smaller models for simpler steps), caching strategies, loop limits to prevent runaway iterations, and explicit cost gates that halt execution when spend thresholds are reached.

Question 7

What is an Agent Control Contract (ACC)?

Accepted Answer

An Agent Control Contract is a structured governance specification that defines an agent's permissions, tool access, autonomy level, escalation policies, and audit requirements. It acts as a declarative policy document that both humans and supervisory agents can enforce at runtime.

Question 8

What are autonomy levels for AI agents?

Accepted Answer

The framework defines four autonomy levels: Assistive (agent suggests, human acts), Delegated (agent proposes a plan, human approves), Bounded Autonomous (agent acts within constraints, escalating edge cases), and Supervisory (agent manages other agents). Each level requires different governance controls.

Question 9

What is an epistemic gate?

Accepted Answer

An epistemic gate is a verification checkpoint in an agent's control loop where the system must demonstrate sufficient confidence or evidence before proceeding. Gates prevent agents from acting on hallucinated or uncertain information by requiring explicit justification at critical decision points.

Question 10

What is harness engineering?

Accepted Answer

Harness engineering is the practice of designing the execution environment and constraints around an agent to improve its performance, rather than only optimising the model itself. Deliberate constraints on token budgets, tool access, and elapsed time have been shown to improve task completion rates by forcing focused, efficient execution.

Frequently Asked Questions