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.

Question 11

What is the AAF MCP server?

Accepted Answer

The AAF MCP (Model Context Protocol) server is a hosted API that exposes the framework as tools for AI assistants such as Cursor, Claude, VS Code, and others. Assistants can look up docs, choose Common Agentic Workloads, analyse trade-offs, load skills, generate Agent Control Contracts, and more — call aaf_guide when unsure which tool to use. After connecting, add the workspace rule from the Tools page so architecture decisions default to AAF. Setup details are on the Tools page.

Question 12

Why is hosted MCP access paid?

Accepted Answer

The framework documentation, GitHub repo, skills, and local CLI remain free. Hosted MCP runs on shared infrastructure with authentication, usage metering, and support. AAF MCP Access is £3/month and includes 1,000 tool calls with a hard monthly cap (no overage). Initialize and tools/list do not count toward the cap. That covers operating the service so the community docs can stay free to read.

Question 13

What do I get with MCP Access, and how do I manage it?

Accepted Answer

You get an API key (shown once after purchase or rotate). Hosted MCP requires Authorization: Bearer <key> on every request to https://www.agenticaf.io/api/mcp — add that header in your IDE MCP config (see the Tools page). Usage is metered as tool calls (1,000/month). Use Manage access to request a magic-link sign-in, see usage, rotate your key, or open the Stripe billing portal. Framework docs, skills in the public GitHub repo, and the local posture CLI do not require a key.

Question 14

Who is WrangleAI?

Accepted Answer

WrangleAI Ltd operates agenticaf.io and the paid AAF MCP Access service (company number 16670149, Malvern, England). WrangleAI builds enterprise AI operations tooling and sponsors the hosted MCP so the Agentic Architecture Framework stays available as open, community-oriented guidance. Stripe Checkout may show WrangleAI as the merchant — that is expected. The framework content itself remains community-driven under its published licence.

Frequently Asked Questions