Back to Blog
Engineering

Scoring Your AI Agent: The 11 Dimensions That Define Production Readiness

Most teams approve agents on subjective review. Here's the structured scoring framework — 11 measurable dimensions across five families — that turns "looks good" into a governance decision you can defend.

'Looks good' is not an evaluation framework. Here are the 11 dimensions every AI agent should be scored on before production — and how each one is actually measured, not guessed at.

August 4, 202614 min read
Agent EvaluationAI GovernanceTrust ScoreProduction ReadinessEnterprise AI
AgentTrustOSAGENT EVALUATIONSCORING · PRODUCTION READINESSScoring Your AI Agent:The 11 Dimensions ThatDefine Production ReadinessCOMPOSITE TRUST SCORE98.2COMPOSITEQuality · Behaviour · SafetyCost / Latency · ExperienceSources: RAGAS · G-Eval Liu et al. 2023 · Berkeley BFCL · PromptBench · AgentTrust OS 2026agent-trust.tech
Composite trust score: one number from 11 dimensions, computed on every governed agent call
Agent Evaluation · Production Readiness · AI Governance
Key Facts
  • RAGAS Faithfulness measures hallucination by decomposing answers into atomic claims and verifying each against retrieved context: score = supported_claims ÷ total_claims. A score below 100 means the agent stated facts it cannot ground in evidence.
  • Tool-Use evaluation matches every tool call against expected name, arguments, and invocation order — with irrelevance detection for calls that should not have been made. Methodology: Berkeley Function Calling Leaderboard.
  • Robustness is tested by perturbing inputs (typo → synonym → paraphrase → semantic shift) and measuring output invariance. An agent that changes its answer when the question is merely rephrased is not production-ready (PromptBench, arXiv 2306.04528).
  • A composite score of 98.2 is achievable with 9 of 11 dimensions scoring automatically from the call envelope — no additional test harness required. Only Faithfulness (requires retrieved context) and Robustness (offline multi-run activity) need extra inputs.
  • Autonomy gates are defined per agent type: a customer-support agent may operate unsupervised at composite ≥ 85 with Safety ≥ 95; a payment agent requires composite ≥ 95 with Faithfulness ≥ 90 plus mandatory human-in-loop regardless of score.
  • G-Eval coherence (Liu et al., 2023) uses a structured LLM-as-judge rubric with chain-of-thought scoring — the most validated approach for dialogue quality evaluation, with higher inter-rater reliability than human-only assessment for long-form outputs.
TL;DR
  • Most teams approve agents for production on "looks good" — a consensus opinion with no measurable criteria and no auditability.
  • 11 dimensions across five families cover every angle of agent trustworthiness: Quality (Accuracy, Faithfulness, Relevance), Behaviour (Coherence, Instruction-Following, Tool-Use), Safety (Safety, Robustness), Cost/Latency, and Experience (Brand & UX).
  • Different dimensions require different measurement methods: deterministic/statistical checks, ML classifiers, or LLM-as-judge — not one approach for everything.
  • A composite score turns "should this agent have more autonomy?" from a meeting into a number — and a governance decision you can defend to your CISO, board, and auditor.
  • Trust Certify sets pre-production pass/fail thresholds per agent type; Trust Runtime scores every live call in under 20ms; Trust Audit produces a tamper-evident record of every evaluation decision.
Keep reading → The 11 dimension breakdown and autonomy gate model below.

At the quarterly AI review, someone asks whether the customer support agent is ready for production. The product manager says the outputs look good. The QA lead says they tested 50 scenarios and found no issues. The CISO says to test more. Nobody can define what "ready" actually means. The meeting ends with a decision to run another two weeks of review.

This is how most enterprise AI agent deployment decisions get made — by subjective consensus, not measurable criteria. There is no defined pass threshold, no scorecard an auditor could examine six months later, and no mechanism to detect whether the agent has drifted from its original quality baseline after deployment. The decision to grant an AI agent more autonomy — to let it execute refunds, close tickets, or schedule follow-ups without a human in the loop — is one of the highest-stakes software decisions a team makes. And it is almost universally made with the same rigor as "let's ship it and see."

This post defines the 11 dimensions every AI agent should be scored on, explains how each is actually measured using validated techniques, and shows how a composite score structure creates the governance mechanism that turns "looks good" into a defensible production decision. We cover the full scoring model, the right measurement technique per dimension, and how Trust Certify, Trust Runtime, and Trust Audit implement this across the agent lifecycle.

Why "looks good" is the wrong evaluation framework for AI agents

Traditional software ships when tests pass. A CI pipeline runs a defined suite; if all green, the deployment proceeds. The threshold is binary, the criteria are objective, and any engineer can reproduce the result. That model breaks completely for AI agents.

An AI agent's behavior is probabilistic, context-dependent, and multidimensional. A response that is accurate might still be incoherent. A response that is coherent might cite a fact the agent fabricated. A tool call that looks right might be invoked with arguments that violate a business policy nobody thought to test. These failure modes do not appear in a binary test suite — they emerge in the interaction of multiple quality signals that need to be evaluated simultaneously.

Key Insight

AI agents fail creatively. A traditional application has a finite set of failure modes. An AI agent has a nearly infinite one. The only way to manage that complexity is to measure it systematically — across every relevant dimension, on every call, with a scoring model that aggregates the signals into a decision.

The second problem with subjective review is that it has no memory. If an agent's behavior drifts after a model update, a prompt revision, or a change in the documents it retrieves, a human reviewer will not notice unless they happen to test the specific scenario where drift occurred. A scoring system running on every call catches drift before a human does — because it is looking at every call, not a curated sample.

The 11 dimensions: five families, every angle covered

The 11 dimensions are organized into five functional families. Each family covers a distinct aspect of agent trustworthiness. None is optional for a production agent — an agent that scores 100 on Quality and 0 on Safety is not a trustworthy agent.

Family 1 — Quality (Dimensions 01–03)
01
Accuracy
Factual correctness of the agent's output against a ground truth reference or retrieved source material.
02
Faithfulness
Every claim in the output is grounded in retrieved context. A faithfulness score below 100 means the agent hallucinated at least one fact.
03
Relevance
The output answers the user's actual question, not a related but different question the agent found easier to answer.
Family 2 — Behaviour (Dimensions 04–06)
04
Coherence
Logical flow and internal consistency across a multi-turn conversation. Drift in tone, contradictions between turns, or broken reference chains all fail coherence.
05
Instruction-Following
Adherence to the system prompt policy and user-specified constraints. If the agent was told not to mention competitors and did, it fails this dimension.
06
Tool-Use
The right tool was called with the right arguments in the right order. Fabricated tool calls, unused results, or out-of-order invocations all fail this dimension.
Family 3 — Safety (Dimensions 07–08)
07
Safety
Harmful content, bias, toxicity, and PII leakage. An agent that surfaces a customer's SSN in a summary has failed Safety regardless of how accurate the summary is.
08
Robustness
Stability under input perturbation. A production agent must give consistent answers when the same question is phrased differently — or it is gaming its own eval suite.
Family 4 — Cost / Latency (Dimensions 09–10)
09
Latency
Time-to-first-token and total response time. An agent that is accurate but takes 12 seconds per call will not survive production at scale.
10
Cost
Dollar-per-task and token efficiency. Agents that solve problems in fewer tokens, without sacrificing quality, have better economics at scale.
Family 5 — Experience (Dimension 11)
11
Brand & UX
Tone, voice, format, and persona fit. An agent that gives technically correct answers in the wrong register — too formal, too casual, wrong terminology for the industry — creates friction and erodes user trust even when it is factually right.

How each dimension is actually measured

This is where evaluation systems divide into those that work and those that don't. The wrong answer is to put a single LLM judge on every dimension and call it done. LLM judges are slow, non-deterministic, expensive per call, and have known biases toward longer outputs and their own prior generations. The right answer is to use the measurement technique that is appropriate for each dimension — deterministic, ML classifier, or LLM-as-judge — and reserve the expensive approach for the cases where it is genuinely necessary.

Faithfulness

Decompose the answer into atomic claims, then verify each claim against retrieved context: score = supported_claims / total_claims. Deterministic baseline uses lexical overlap; LLM-upgradeable to full RAGAS semantic verification.

Relevance

Generate the questions the answer would answer, then measure cosine similarity to the actual question. An answer that would answer a different question is penalized — even if it is factually accurate and well-written.

Coherence

LLM-as-judge with a chain-of-thought rubric. The judge scores logical flow, internal consistency, and narrative continuity across turns. Requires an LLM judge but produces reliable, human-interpretable explanations for failures.

Tool-Use

Match tool calls against expected name, arguments, and invocation order. Includes irrelevance detection — a tool call that should not have been made is a failure even if it returned a valid result.

Robustness

Input perturbation: typo → synonym → paraphrase → semantic shift. Score = output invariance across perturbation levels. This is an offline, multi-run evaluation — it cannot be done from a single call, but must be run against every prompt change or model update.

Safety

Three-layer detection: toxicity classifiers (Detoxify), PII detectors (Presidio for SSNs, API keys, payment card numbers via Luhn validation), and safety-judge models (Llama Guard for harmful content). Each layer catches a different category of safety failure.

Source: Detoxify · Microsoft Presidio · Meta Llama Guard

The underlying pattern: some dimensions are deterministic or statistical, some require ML classifiers, and some genuinely need an LLM judge. A governance-grade evaluation layer uses the right method for each — not one LLM for everything. Using an LLM to evaluate latency is wasteful. Using a string match to evaluate coherence misses the point.

● Deterministic / statistical — Tool-Use, Latency, Cost, PII● ML classifier — Safety / Toxicity, Robustness● LLM-as-judge — Coherence, Brand & UX, deep Faithfulness
11
Dimensions across five families — the complete evaluation surface for any production AI agent
Quality · Behaviour · Safety · Cost · Experience
9/11
Dimensions that score automatically from the call envelope — no extra test harness required
Faithfulness and Robustness require additional inputs
98.2
Composite score threshold used by high-reliability agents (payment processing, medical records)
AgentTrust OS governance model, 2026

The composite score: from dimensions to an autonomy gate

Individual dimension scores are useful for debugging — they tell you exactly which aspect of agent behavior failed and why. The composite score is useful for governance — it tells you whether the agent clears the bar to operate at a given level of autonomy.

Each dimension carries a weight that reflects its importance for the agent's specific use case and risk context. Safety and Faithfulness are weighted higher for agents that handle sensitive data or make consequential decisions. Latency and Cost are weighted higher for agents in high-volume, cost-sensitive pipelines. The resulting composite is a single number from 0 to 100 that encodes the full quality surface of one agent execution.

Pro Tip

Don't use a single composite threshold across all your agents. A customer support agent and a payment execution agent have fundamentally different risk profiles. Define per-agent-type thresholds: what score enables unsupervised operation, what score triggers human review, and what score forces a block regardless of business pressure.

The composite score gates autonomy promotion. This is the structural link between measurement and governance: you cannot grant an agent more autonomy unless it has demonstrated it meets the measurable criteria for that autonomy level. This transforms the deployment review from "does this feel ready?" to "does it meet the defined threshold?"

Customer Support Agent
Composite ≥ 85
Safety ≥ 95
→ Runs unsupervised
Payment Execution Agent
Composite ≥ 95
Faithfulness ≥ 90
Human-in-loop required

Note that the payment agent requires both a high composite and a specific Faithfulness floor — and mandates human-in-loop regardless of score. This is the correct model for high-stakes agents: the composite score governs routine operations, while explicit hard constraints handle the highest-consequence scenarios where no score, however high, eliminates the need for human review.

11 Dimensions → Composite Score → Autonomy GateQUALITYAccuracy · FaithfulnessRelevanceBEHAVIOURCoherence · InstructionTool-UseSAFETYSafety · RobustnessCOST / LATENCYLatency · CostEXPERIENCEBrand & UXCOMPOSITE98.2/ 100AUTONOMY GATE≥ threshold?APPROVEESCALATEBLOCK
Figure 1: The 11-dimension evaluation model. Five families feed into a composite score, which gates the autonomy decision: approve for unsupervised execution, escalate for human review, or block when thresholds are not met.

Deterministic-first, LLM-upgradeable: the only viable architecture

A common mistake is building an evaluation layer that runs an LLM judge on every request. In testing, this looks thorough. In production, it collapses. LLM judges add 800ms–5 seconds of latency per call, cost money on every inference, produce non-deterministic outputs across identical inputs, and cannot explain their reasoning in an auditable way. When an LLM judge says a response "feels off," that is not a reason you can put in an incident report or regulatory filing.

The viable architecture is deterministic-first, LLM-upgradeable. The baseline implementation uses statistical and rule-based methods for each dimension — 1–5ms per check, free, reproducible, and fully explainable. For dimensions where LLM-grade precision is necessary (Faithfulness in a RAG pipeline, Brand & UX for a high-stakes customer-facing agent), you register an LLM judge that upgrades that specific dimension without changing anything else. The result: 9 of 11 dimensions resolve in under 5ms total, LLM judges run only on the 1–2 dimensions where they add genuine value, and every failure produces a concrete, auditable reason — the exact claim that could not be grounded, the exact tool that was called out of order, the exact PII pattern that was detected.

What does a real failure look like in practice? Here is a decision log from four live agent calls:

Decision Log — Trust Runtime
✓ faq-agent → APPROVE   composite=100   faithfulness=100   safety=100
✖ support-agent → BLOCK   policy=0   reason: data leakage: US Social Security Number
✖ chat-agent → BLOCK   policy=20   reason: prompt injection: instruction override attempt
⚠ payment-agent → ESCALATE   composite=91   faithfulness=82 → below threshold (90)

The clean answer sails through in under 20ms. The SSN leak, the jailbreak attempt, and the low-faithfulness payment response are each stopped with a specific, reproducible reason. No "the model felt off." Every decision can be audited, replicated, and explained to a regulator.

Pro Tip

When running Robustness evaluation offline, test every prompt template change — not just model updates. A prompt that routes differently under paraphrase is as broken as a model that does. Most teams test model updates and forget that prompt engineering is equally capable of introducing robustness regressions.

The DX problem: scoring in production, not just in a test harness

Most evaluation frameworks — RAGAS, DeepEval, TruLens, Galileo — are designed for offline benchmark runs. You curate a test set, configure the evaluation pipeline, run it against a snapshot of your model and prompts, and get a report. This is genuinely useful for pre-deployment validation and regression detection between model versions.

What these frameworks do not do: score every production call, return a gate decision synchronously before the output is released, or produce audit records as a byproduct of normal operation. They are designed for evaluation cycles, not governance infrastructure. The distinction matters because agent failures do not happen in your test set — they happen in production, with real users, in scenarios your test set never covered.

Key Insight

Eval frameworks and runtime governance are complementary, not competitive. Use eval frameworks for comprehensive pre-deployment validation and regression testing between model versions. Use runtime governance for inline scoring of every production call, synchronous gate decisions, and audit records that regulators and boards can inspect. The failure to distinguish these two jobs is why teams think they have evaluation covered when they actually only have the offline half of it.

The practical implication for teams: zero-config scoring of 9 of 11 dimensions from the call envelope means there is no reason not to score every production call. The call already carries the agent ID, request, tool execution trace, output, latency, and token count — 9 dimensions score from that data with no additional instrumentation. Faithfulness needs the retrieved context; Robustness runs offline. Everything else is already there, already structured, already available. The only choice is whether to compute the scores or discard that information.

How to set initial thresholds — and when to adjust them

The question every team asks after seeing the 11-dimension model is: where do I start? The honest answer is that your initial thresholds are educated guesses that you will adjust based on production data. Start conservative, measure the false positive rate (valid outputs that get escalated or blocked), and tighten or loosen specific dimension floors based on what the data shows.

A practical starting framework for a customer-facing support agent: Safety ≥ 95 (hard floor — do not negotiate this), Faithfulness ≥ 80 (escalate for human review below this), Tool-Use ≥ 90 (block on fabricated tool calls regardless of composite), composite ≥ 80 for unsupervised operation. Run this in shadow mode for two weeks — compute every decision but don't block — and review the escalation queue. If 40% of escalations are false positives (human reviewers approve them), raise the Faithfulness floor. If Safety escalations are all genuine, keep that floor or raise it.

For high-stakes agents (payment, records modification, scheduling with external consequences), start with composite ≥ 95, Safety ≥ 99, Faithfulness ≥ 90, and mandatory human-in-loop for any action above a defined impact threshold regardless of score. Build toward autonomous operation by accumulating evidence, not by lowering the bar.

Monitoring composite score drift after deployment

Production readiness is not a one-time certification. An agent that scores 95 on launch day may score 82 six weeks later — because the documents it retrieves have changed, a new user population is probing it in ways that weren't in the test set, or an upstream data source has degraded. Score drift is real, it is common, and it is invisible without continuous monitoring.

Track composite score as a time-series metric, not just a one-time gate. Set alert thresholds on rolling averages: if the 7-day average composite drops below the certification score by more than 5 points, trigger a re-evaluation. Track dimension-level drift separately — a drop in Faithfulness with stable Safety and Coherence usually points to a retrieval degradation problem, not a model problem. A drop in Tool-Use often signals that an API schema changed and the agent is now calling tools with stale argument formats.

AgentTrust OS: The Governance Platform for 11-Dimension Agent Scoring

AgentTrust OS implements the 11-dimension evaluation framework across the full agent lifecycle — from pre-production certification through every live call to long-term audit reporting.

Trust Certify
Pre-production: defines per-agent-type thresholds across all 11 dimensions, runs the full evaluation suite including offline Robustness testing, and certifies agents for their designated autonomy tier before they touch production traffic.
Trust Runtime
Live scoring: evaluates every governed call against all 11 dimensions in under 20ms, computes the composite score, and returns an approve / escalate / block decision with the exact dimension and rule that determined the outcome.
Trust Audit
Governance record: every evaluation — dimension scores, composite, decision, and the specific rule that fired — is stored as a tamper-evident audit record. Six months later, you can reconstruct exactly what the agent did, what it scored, and why it was allowed to proceed.
Trust CertifyPre-production certification11-dim thresholds definedTrust Runtime11-dim scoring per callComposite → gate decisionTrust AuditTamper-evident score recordAuditable decision history
Figure 2: AgentTrust OS governance pipeline — from pre-production threshold definition through live per-call scoring to long-term auditable record.

Frequently Asked Questions

Nine of the eleven dimension scores can be computed automatically from the call envelope — agent ID, request, tool execution trace, and output — with no additional inputs. Faithfulness requires the retrieved context passed alongside the output; Robustness is an offline, multi-run evaluation that runs against prompt templates and model versions, not on each live call. For high-throughput pipelines, you can configure dimension-specific sampling rates: run the full 11-dimension suite on a defined percentage of calls and the lightweight deterministic subset on every call.
Standalone eval frameworks run offline, in a test harness, against curated test sets. They are excellent for catching regressions between model versions and for pre-deployment validation. What they don't do is evaluate production calls in real time, return a gate decision synchronously, or produce audit records as a byproduct of every live request. The 11-dimension model is compatible with RAGAS and G-Eval methodologies — it uses them where they are the right tool — but implements them as an inline governance layer rather than an offline evaluation pipeline. The two are complementary: use eval frameworks for comprehensive pre-deployment testing, Trust Runtime for live governance of every production call.
There is no universal number. The threshold should be determined by the agent's use case, its blast radius (the scope of systems and actions it can affect), and the reversibility of its actions. A customer support agent that drafts email replies for human review can operate autonomously at composite ≥ 85 with minimal risk. An agent that executes payments, modifies customer records, or makes credit decisions should require composite ≥ 95 with explicit per-dimension floors on Safety and Faithfulness — and may still require human-in-loop regardless of score for the highest-consequence action classes. Start with a conservative threshold, gather data on failure rate and false positive rate, then adjust upward with evidence.
Each agent in a multi-agent workflow should be scored independently at its output boundary. The composite score of a downstream agent does not inherit the quality of its upstream inputs — if an upstream agent passes hallucinated context, the downstream agent's Faithfulness score will reflect that. This is a feature, not a limitation: it means governance catches degraded data propagation between agents rather than laundering a quality failure through the pipeline. For orchestration layers, aggregate scoring across the full workflow provides the composite trust signal for the end-to-end operation.
It is honest, not aspirational — but the quality of measurement varies by dimension. Latency, Cost, Tool-Use (deterministic), and Safety (with established classifiers) are measured with high precision today. Accuracy depends on having a ground truth reference, which you may not always have in production. Robustness requires offline test infrastructure. Brand & UX requires a configured rubric and an LLM judge that matches your quality bar. The framework is designed to be progressively upgradeable: deterministic proxies work out of the box, and each dimension improves as you add more inputs and configure more sophisticated judges.
Ready to Score Your Agents?

From "Looks Good" to a Defensible Number

AgentTrust OS certifies agents on all 11 dimensions — pre-production, at runtime, and across every audit record.

Start Free →

More from the blog

AI ComplianceJuly 22, 2026AI ComplianceJuly 22, 2026AI GovernanceJuly 22, 2026AI GovernanceJuly 22, 2026AI ArchitectureJuly 22, 2026AI SecurityJuly 16, 2026MLOpsJuly 10, 2026AI ImplementationJuly 8, 2026AI Agent ArchitectureJuly 5, 2026AI Agent ArchitectureJune 30, 2026EngineeringJune 23, 2026AI Agent ArchitectureJuly 2, 2026AI Agent ArchitectureJuly 1, 2026IntegrationsJuly 1, 2026IntegrationsJuly 1, 2026IntegrationsJuly 2, 2026AI SecurityJuly 3, 2026AI SecurityJuly 1, 2026AI ComplianceJuly 2, 2026AI ComplianceJuly 3, 2026AI StrategyJuly 2, 2026AI StrategyJuly 3, 2026AI StrategyJuly 3, 2026AI StrategyJuly 3, 2026AI GovernanceJuly 28, 2026Healthcare AIJuly 28, 2026ArchitectureJuly 29, 2026ArchitectureJuly 29, 2026AI StrategyJuly 29, 2026AI StrategyJuly 30, 2026AI SecurityJuly 30, 2026AI ComplianceJuly 30, 2026EngineeringJuly 30, 2026AI GovernanceAugust 4, 2026EngineeringAugust 4, 2026