The 8 Dimensions of AI Agent Intelligence

Most teams think of their AI agent as a single thing: "It's a support bot" or "It's a sales assistant." This is like describing a car as "transportation" — technically true, but useless for diagnosing problems.

When an agent fails, the failure always traces back to one of eight specific dimensions. Understanding which dimension failed tells you exactly what to fix. A guardrail failure needs a different fix than a domain knowledge gap, which needs a different fix than a scope boundary violation.

Agent Scrimmage's Discovery Engine maps all 8 dimensions in about 30 seconds. This creates the ground truth for every test that follows.

Capabilities are what your agent can do. Process refunds. Look up orders. Schedule appointments. Send emails. Create tickets. Query a CRM.

The gap between claimed capabilities and actual capabilities is where most failures hide. An agent might claim it can "manage customer accounts" — but can it actually update a shipping address? Reset a password? Cancel a subscription? Each sub-capability needs independent verification.

Discovery extracts capabilities by analyzing skill files (what the agent is configured to do) and probing with diagnostic messages (what it actually does). From our first production test, we extracted 42 confirmed capabilities in under 30 seconds.

What to test: Can the agent perform each claimed capability correctly? Does it handle partial capabilities (can look up orders but can't modify them)? Does it fail gracefully when a capability is unavailable?

Guardrails are hard rules the agent must follow. "Never process refunds over $500 without manager approval." "Always escalate legal threats to compliance." "Never share internal pricing formulas."

Guardrails are the most critical dimension because when they fail, the consequences are immediate and severe — unauthorized refunds, compliance violations, data leaks.

The problem: guardrails are easy to define but hard to enforce. An agent can have a clear rule saying "never approve refunds over $500" and still approve a $4,200 refund when a customer combines emotional pressure with a legal threat. The guardrail exists in the configuration, but the agent's behavior under pressure overrides it.

What to test: Does each guardrail hold under adversarial pressure? Can the guardrail be bypassed through social engineering, prompt injection, or context manipulation? Does the agent recognize when a guardrail applies in ambiguous situations?

Limitations are what the agent explicitly cannot do. "Cannot access the billing system." "Cannot make outbound phone calls." "Cannot modify user permissions."

A well-designed agent acknowledges its limitations honestly. A poorly-designed agent hallucinates — it pretends it can do something it can't, or it gives a plausible-sounding but completely fabricated answer.

The most dangerous limitation failures are "soft hallucinations" — the agent doesn't make up facts, but it implies capabilities it doesn't have. "I can look into that for you" when it actually has no access to the system in question. The customer waits, the agent stalls, and trust erodes.

What to test: Does the agent decline gracefully when asked to do something it can't? Does it hallucinate capabilities? Does it clearly communicate what it can't do and offer alternatives?

Tools are the external systems the agent connects to: Salesforce, HubSpot, Stripe, PostgreSQL, file systems, APIs, MCP servers. Every tool connection is a potential failure point.

Tool failures are infrastructure problems, not conversation problems. The agent might have a perfect response ready, but if the API call fails, the database is slow, or the CRM returns stale data, the user gets a broken experience.

Agent Scrimmage maps every tool connection during discovery and simulates realistic tool interactions during scenario testing. For skill-file agents, we provide mock infrastructure — simulated CRM data, file systems, and API endpoints — so the agent can demonstrate its full workflow without touching production.

What to test: Does the agent handle tool failures gracefully? Does it retry, fall back, or inform the user? Does it use the right tool for each task? Does it handle stale or inconsistent data from tools?

Domain knowledge is industry-specific expertise baked into the agent. HIPAA rules for healthcare. FMCSA regulations for logistics. Fair Housing Act for real estate. FDA Food Code for hospitality. Tax thresholds for finance.

Domain knowledge failures are the hardest to detect because the agent sounds confident. It gives a well-structured, professional answer that happens to be wrong. A healthcare agent that cites a HIPAA rule that doesn't exist. A logistics agent that states an incorrect weight limit. A real estate agent that misquotes Fair Housing requirements.

Agent Scrimmage's scenarios are grounded in real regulatory data — actual regulations, actual industry metrics, documented compliance requirements. We verify domain knowledge accuracy against real-world sources, not just whether the answer sounds right.

What to test: Are the agent's domain-specific claims factually accurate? Does it cite regulations correctly? Does it acknowledge uncertainty when it's not sure? Does it stay current with regulatory changes?

Persona is how the agent communicates — its tone, style, empathy level, and escalation behavior. Professional vs. casual. Empathetic vs. direct. Formal vs. conversational.

Persona consistency under pressure is the real test. Any agent can be professional when asked "What are your business hours?" The question is whether it stays professional when a customer says "This is the worst company I've ever dealt with and I want a full refund RIGHT NOW."

The most common persona failure is sycophancy — the agent becomes so eager to please that it agrees with everything, makes promises it can't keep, or validates complaints that aren't justified. The opposite failure is coldness — opening with a scope disclaimer instead of empathy when a customer is in distress.

What to test: Does the agent maintain persona under emotional pressure? Does it avoid sycophancy? Does it lead with empathy before scope disclaimers? Does it match the appropriate tone for each situation?

Data access is what the agent can read and write. Customer records. Order history. Internal documents. Knowledge bases. Persistent memory across conversations.

Data access defines the agent's "view of the world." An agent with access to CRM records can personalize responses. An agent without that access is generic. But data access also creates risk — an agent that can read customer records can potentially leak them.

The most common data access failure is the "context loss" problem. The agent has access to data from earlier in the conversation, but loses it due to context window limits or technical errors. It then denies the prior conversation existed, frustrating the customer and destroying trust.

What to test: Can the agent access the data it needs? Does it handle missing or stale data? Does it maintain context across conversation turns? Does it protect sensitive data from unauthorized disclosure?

Boundaries define where the agent's scope starts and ends. "In scope: billing questions, order status, refund requests." "Out of scope: technical support, product development, legal advice."

Boundary failures come in two forms:

1. 1.The agent attempts an out-of-scope request and gives a wrong or harmful answer (overreach)
2. 2.The agent refuses an in-scope request because it's unsure whether it falls within its scope (underreach)

Both are costly. Overreach creates liability. Underreach frustrates customers who have legitimate requests.

The hardest boundary cases are "gray zone" requests — questions that are adjacent to the agent's scope but not explicitly included or excluded. "Can you recommend a lawyer?" for a support agent. "What's the weather like?" for a scheduling agent. These require judgment, not just rule-following.

What to test: Does the agent correctly identify in-scope, boundary, and out-of-scope requests? Does it deflect out-of-scope requests gracefully? Does it handle gray zone requests with appropriate judgment? Does it refer to the right resource when declining?

Every AI agent failure traces to one or more of these 8 dimensions. When you map all 8 before testing, you know exactly what to test and exactly what to fix when something breaks.

Agent Scrimmage's Discovery Engine maps all 8 dimensions in about 30 seconds — extracting capabilities, guardrails, limitations, tools, domain knowledge, persona patterns, data access, and scope boundaries from your agent's skill files or API responses.

Then scenario simulations test each dimension systematically. Not generic prompts — targeted scenarios designed to stress the specific dimensions your agent has. The result is a readiness score, failure analysis, and training assets to fix every gap we find.

Discovery is free. Connect your agent and see all 8 dimensions mapped in under a minute.