Agentic – Evidence & Audit

Overview: Evidence in the Agentic Context

Evidence is the proof that an Agentic control is not only documented but actually implemented and effective. WAF++ distinguishes two classes:

Required Evidence: Must exist for every control; absence is an audit finding
Optional Evidence: Recommended for higher maturity levels and regulated environments

Evidence should ideally be generated automatically — by the WAF++ Checker, Terraform outputs, agent tracing systems, CloudWatch exports, or SIEM reports.

Evidence Types for Agentic Systems

Evidence Type	Description	Typical Sources
Agent Telemetry	Token consumption, latency, error rates, GPU/CPU utilization	CloudWatch, Prometheus, Application Insights
Approval Logs	Human-in-the-loop decisions with timestamp and approver identity	Workflow system, SIEM
Safety Checks	Guardrail test results and alignment checks	Bedrock Guardrails, Vertex AI Safety, Azure Content Filters
Model Cards	Documented model capabilities, limitations, and usage risks	ML model registry, wiki
Reasoning Traces	Recordings of chain-of-thought, tool calls, and decision paths	Object storage, tracing backend

Evidence Type

Description

Typical Sources

Agent Telemetry

Token consumption, latency, error rates, GPU/CPU utilization

CloudWatch, Prometheus, Application Insights

Approval Logs

Human-in-the-loop decisions with timestamp and approver identity

Workflow system, SIEM

Safety Checks

Guardrail test results and alignment checks

Bedrock Guardrails, Vertex AI Safety, Azure Content Filters

Model Cards

Documented model capabilities, limitations, and usage risks

ML model registry, wiki

Reasoning Traces

Recordings of chain-of-thought, tool calls, and decision paths

Object storage, tracing backend

Evidence Table: All WAF-AGN Controls

Control	Required Evidence	Format
WAF-AGN-010	Agent inventory; stable IDs; auth bindings; interaction audit logs	HCL, JSON logs, Markdown register
WAF-AGN-020	Tool whitelist/schemas; IAM policies; tool-use audit logs with context	HCL, JSON/SIEM export
WAF-AGN-030	Inference logging; chain-of-thought; tool-call details; trace analysis	Terraform, JSON export, dashboard
WAF-AGN-040	Versioned memory storage; backup/PITR; audit logs of state changes	HCL, backup test reports
WAF-AGN-050	Guardrail configuration; content-filter policies; action boundaries; HIL workflows	HCL, policy files, test reports
WAF-AGN-060	Coordinator agent; event channels; retry/DLQ; health monitoring dashboards	HCL, Terraform export, runbook
WAF-AGN-070	Token/latency/error metrics; X-Ray/OpenTelemetry tracing; alerting	HCL, dashboard JSON, metric export
WAF-AGN-080	HIL workflow; approver lists; complete approval audit logs	HCL, JSON/SIEM, runbook
WAF-AGN-090	Budget/alarm configuration; cost analysis; cache configuration	HCL, spreadsheet, review record
WAF-AGN-100	Agent Debt Register with limitation, impact, owner, target date; review minutes	Markdown/database export, PDF

Control

Required Evidence

Format

WAF-AGN-010

Agent inventory; stable IDs; auth bindings; interaction audit logs

HCL, JSON logs, Markdown register

WAF-AGN-020

Tool whitelist/schemas; IAM policies; tool-use audit logs with context

HCL, JSON/SIEM export

WAF-AGN-030

Inference logging; chain-of-thought; tool-call details; trace analysis

Terraform, JSON export, dashboard

WAF-AGN-040

Versioned memory storage; backup/PITR; audit logs of state changes

HCL, backup test reports

WAF-AGN-050

Guardrail configuration; content-filter policies; action boundaries; HIL workflows

HCL, policy files, test reports

WAF-AGN-060

Coordinator agent; event channels; retry/DLQ; health monitoring dashboards

HCL, Terraform export, runbook

WAF-AGN-070

Token/latency/error metrics; X-Ray/OpenTelemetry tracing; alerting

HCL, dashboard JSON, metric export

WAF-AGN-080

HIL workflow; approver lists; complete approval audit logs

HCL, JSON/SIEM, runbook

WAF-AGN-090

Budget/alarm configuration; cost analysis; cache configuration

HCL, spreadsheet, review record

WAF-AGN-100

Agent Debt Register with limitation, impact, owner, target date; review minutes

Markdown/database export, PDF

Evidence Retention Periods

Evidence Type	Minimum Retention	Requirement	Recommendation
Agent Interaction Logs	1 year	SOC 2, ISO 27001, DORA	1 year hot, 3 years cold (Glacier)
Reasoning Traces	1 year	SOC 2, GDPR Art. 22	Object storage with versioning
Tool-Use Audit Logs	1 year	SOC 2 CC7, ISO 27001 A.8.2	SIEM archive with immutable storage
Human-in-the-Loop Approvals	3 years	ISO 27001 A.8.2, GDPR	Workflow platform or Git
Agent Telemetry	90 days	SOC 2 CC4/CC7	Monitoring with downsampling
Guardrail Test Reports	1 year	ISO 27001 A.5.36, NIST CSF	S3 archive with lifecycle policy
Agent Debt Register	3 years	ISO 27001 A.5.33, DORA	Git or ITSM
Model Cards	Lifetime + 3 years	EU AI Act, ISO 27001	ML registry with versioning

Evidence Type

Minimum Retention

Requirement

Recommendation

Agent Interaction Logs

1 year

SOC 2, ISO 27001, DORA

1 year hot, 3 years cold (Glacier)

Reasoning Traces

1 year

SOC 2, GDPR Art. 22

Object storage with versioning

Tool-Use Audit Logs

1 year

SOC 2 CC7, ISO 27001 A.8.2

SIEM archive with immutable storage

Human-in-the-Loop Approvals

3 years

ISO 27001 A.8.2, GDPR

Workflow platform or Git

Agent Telemetry

90 days

SOC 2 CC4/CC7

Monitoring with downsampling

Guardrail Test Reports

1 year

ISO 27001 A.5.36, NIST CSF

S3 archive with lifecycle policy

Agent Debt Register

3 years

ISO 27001 A.5.33, DORA

Git or ITSM

Model Cards

Lifetime + 3 years

EU AI Act, ISO 27001

ML registry with versioning

Audit Preparation Checklist

Before the Audit (4–6 Weeks Out)

Run the WAF++ Checker against all Agentic IaC and address findings
Verify completeness of agent interaction logs (agent ID, timestamp, outcome)
Review tool-use logs from the last 90 days for approval and compliance
Generate guardrail test reports for the last 90 days and document outliers
Verify HIL approvals include full context and valid approvers
Check the Agent Debt Register for currency and closed improvement plans

Evidence Collection

Export Terraform code and document the commit hash
Export WAF++ Checker report as PDF
Export agent logs for the audit period as JSON/CSV
Archive reasoning traces for critical decisions
Extract tool-use audit logs from SIEM or object storage
Create HIL approval export for sensitive decisions
Compile guardrail test and alignment-check results

Common Auditor Questions and WAF++ Answers

Auditor Question	WAF++ Evidence
"How do you ensure every agent is uniquely identifiable?"	Agent inventory (WAF-AGN-010); Terraform IDs; IAM/service account bindings
"How do you prevent unauthorized tool use?"	Tool whitelist and IAM policies (WAF-AGN-020); tool-use audit logs; WAF++ Checker gate
"How do you explain agent decisions?"	Reasoning traces with chain-of-thought (WAF-AGN-030); immutable audit trail
"How do you protect memory and state?"	Versioned storage, backups, point-in-time recovery (WAF-AGN-040)
"How do you ensure human review of sensitive decisions?"	HIL workflow with audit trail (WAF-AGN-080); approver lists and emergency intervention

Auditor Question

WAF++ Evidence

"How do you ensure every agent is uniquely identifiable?"

Agent inventory (WAF-AGN-010); Terraform IDs; IAM/service account bindings

"How do you prevent unauthorized tool use?"

Tool whitelist and IAM policies (WAF-AGN-020); tool-use audit logs; WAF++ Checker gate

"How do you explain agent decisions?"

Reasoning traces with chain-of-thought (WAF-AGN-030); immutable audit trail

"How do you protect memory and state?"

Versioned storage, backups, point-in-time recovery (WAF-AGN-040)

"How do you ensure human review of sensitive decisions?"

HIL workflow with audit trail (WAF-AGN-080); approver lists and emergency intervention

Generate Evidence from wafpass

The WAF++ Checker (wafpass) can export evidence directly for audits:

# Check all Agentic controls and generate a PDF report
wafpass check \
  --pillar agentic \
  --path ./infrastructure \
  --format pdf \
  --output audit-evidence/wafpass-agentic-$(date +%Y%m%d).pdf

# Export reasoning traces as machine-readable evidence
wafpass check \
  --pillar agentic \
  --control WAF-AGN-030 \
  --path ./infrastructure \
  --format json \
  --output audit-evidence/agentic-reasoning-$(date +%Y%m%d).json

# Check a single control
wafpass check \
  --control WAF-AGN-050 \
  --path ./infrastructure \
  --format github-annotations

Common Audit Findings and WAF++ Remediations

Audit Finding	Control	WAF++ Remediation
Agents without unique identity or auth	WAF-AGN-010	Stable IDs in Terraform; IAM/service account bindings; inventory register
Unrestricted tool use	WAF-AGN-020	Tool whitelist and schemas; IAM policies; tool-use logging
Decisions not traceable	WAF-AGN-030	Inference logging; chain-of-thought traces; immutable storage
No human approval for critical actions	WAF-AGN-080	HIL workflow; approval audit trail; emergency intervention procedure
Missing Agent Debt Register	WAF-AGN-100	Quarterly reviews; issue-tracking integration; improvement plans

Audit Finding

Control

WAF++ Remediation

Agents without unique identity or auth

WAF-AGN-010

Stable IDs in Terraform; IAM/service account bindings; inventory register

Unrestricted tool use

WAF-AGN-020

Tool whitelist and schemas; IAM policies; tool-use logging

Decisions not traceable

WAF-AGN-030

Inference logging; chain-of-thought traces; immutable storage

No human approval for critical actions

WAF-AGN-080

HIL workflow; approval audit trail; emergency intervention procedure

Missing Agent Debt Register

WAF-AGN-100

Quarterly reviews; issue-tracking integration; improvement plans