Continuous AI agent evaluation measures the full production workflow, not just the model. VeriRFP runs recurring benchmark checks so teams can see drift before buyers do.
When an enterprise buyer asks "how do you ensure your AI is accurate?" the honest answer from most platforms is some version of "we tested it before deployment." That answer is no longer sufficient.
AI agents drift. Models update. Evidence libraries grow. Retrieval patterns change. An agent that performed well during initial testing can silently degrade in production without an obvious failure mode. The confidence scores stay high. The outputs still look reasonable. Accuracy erodes anyway.
VeriRFP now continuously evaluates every AI agent against golden benchmarks on a recurring cadence. This is not model evaluation alone. It tests the full path from input through retrieval, reasoning, evidence citation, and final output.
TL;DR
- VeriRFP evaluates every production agent against golden examples every six hours on Business and Enterprise plans.
- The system scores response accuracy, evidence citation, hallucination risk, tool use, safety, latency, and confidence calibration.
- A score drop greater than 10% triggers an anomaly alert for review.
- The AI Governance Dashboard turns those scores into audit evidence for security and procurement teams.
Why Agent Evaluation Differs From Model Evaluation
Agent evaluation matters because buyers approve workflows, not raw models.
Model evaluation tests a language model's raw capability. It checks whether the model can answer questions, classify text, or generate coherent prose. Benchmarks like MMLU and HumanEval measure the model in isolation.
Agent evaluation tests the composed system. It checks the retrieval pipeline, the reasoning chain, the output format, and the confidence score that calibrates trust. A model upgrade that improves language quality can still hurt retrieval precision if the embedding space shifts.
The distinction matters because enterprise buyers are not purchasing a language model. They are purchasing a workflow that uses their evidence and produces answers their security team will approve.
Seven Scoring Criteria
Each criterion maps to a specific production risk.
VeriRFP evaluates agents across seven criteria, each targeting a different dimension of quality:
Response Accuracy
Response accuracy checks whether the output matches the expected response. For structured fields, it verifies key presence and value matching. For text fields, it uses normalized similarity scoring. Different phrasing can still score well. Missing details or incorrect claims score poorly.
Evidence Citation
Evidence citation checks whether the answer is backed by approved sources. It verifies that the agent references specific evidence documents, uses relevant citations, and supports its claims with traceable sources.
Hallucination Detection
Hallucination detection checks whether the output contains unsupported claims. It compares response statements against the approved evidence corpus. Any claim that cannot be traced to a source is flagged. The threshold is strict. One unsupported claim in ten outputs triggers a warning.
Tool Trajectory
Tool trajectory checks whether the agent used the right tools in the right order. Multi-step agents follow a planned sequence: retrieve evidence, evaluate adequacy, draft the response, then score confidence. Deviations often signal retrieval failures or reasoning shortcuts.
Safety
Safety checks whether the output exposes PII, secrets, or injection patterns. It runs the same detection patterns used in the live security pipeline. This catches cases where the agent surfaces sensitive data from the evidence corpus.
Latency Budget
Latency budget checks whether the agent completed within the expected time bound. Enterprise workflows still need a usable response time. A correct answer in 30 seconds fails if the budget is 10 seconds. Scoring decays linearly from the budget threshold to zero at three times the limit.
Confidence Calibration
Confidence calibration checks whether the agent correctly estimates its own reliability. An agent that reports 85% confidence should be correct about 85% of the time. An agent that reports 95% confidence but is correct only 70% of the time is overconfident. This criterion measures that gap.
Golden Evaluation Sets
Golden evaluation sets give each agent a repeatable baseline.
Each agent type has a curated set of realistic inputs with known-correct outputs. VeriRFP derives these examples from representative production scenarios.
| Agent Type | Examples | Primary Criteria |
|---|---|---|
| Draft Pipeline | 45 | Response accuracy, evidence citation, hallucination |
| Compliance Heartbeat | 30 | Drift detection accuracy, false positive rate |
| Smart Requestionnaire | 40 | Question classification, routing accuracy |
| Retrieval Evaluator | 40 | Retrieval precision, relevance scoring |
| Quality Check | 35 | Quality gate accuracy, consistency |
The evaluation engine runs every six hours for workspaces on Business and Enterprise plans. Score declines greater than 10% from the previous evaluation trigger automated anomaly alerts.
The AI Governance Dashboard
The dashboard turns evaluation runs into evidence that teams can inspect.
Evaluation scores feed into the AI Governance Dashboard, available on Business and Enterprise workspaces. The dashboard provides:
- Agent health cards showing current autonomy level, circuit breaker state, and latest evaluation score per agent type
- Quality trend charts tracking scores over 30 days so teams can spot gradual degradation
- Anomaly alert feed with severity classification (info, warning, critical) and resolution tracking
- ATF compliance status organized by the five Agentic Trust Framework elements
- Searchable audit trail of recent agent executions with confidence scores and durations
This is not a vanity dashboard. Security teams evaluating VeriRFP during buyer diligence can review it to confirm that AI agents stay within governed boundaries. The evaluation data shows that the governance claims on the security page are backed by runtime measurement.
What Comes Next
Continuous evaluation is the gate for higher autonomy.
As agents demonstrate consistent quality across evaluation cycles, they become eligible for promotion to higher autonomy levels in the ATF maturity model. Agents that maintain Level 3 (Senior) quality thresholds across 30 days of evaluation become candidates for Level 4 (Principal) promotion.
This creates a virtuous cycle: governance enables trust, trust enables autonomy, autonomy enables efficiency, and continuous evaluation ensures the cycle holds.
The full ATF conformance matrix and governance architecture are documented at verirfp.com/ai-governance. The AI Governance Dashboard is included in Business and Enterprise plans.