- Home
- Skills
- Agents & Orchestration
- Agentflow Auditor
Works with the AI Agents you already use
Agentflow Auditor
Audit AI agent architectures for security, reliability, and delegation safety using evidence-based gap analysis.
$9.99
Agentflow Auditor
Example session with this skill installed
Audit these agent orchestrator files for missing guardrails and give me a prioritized remediation plan based on the rubric. I have attached the tool gateway config and the delegation logic.
- Read your context and instructions
- Compiled the agentflow auditor
AgentFlow Auditor — Sample Audit Report
Demonstration status: Illustrative example
Testing context: AgentFlow Auditor has undergone ChatGPT-assisted development and local automated testing. This sample demonstrates the intended audit format; it is not evidence of a live repository audit or production security certification.
1. Executive Summary
AgentFlow Auditor has assessed a fictional multi-agent orchestration system consisting of a CoordinatorAgent, ToolExecutor, tool gateway and supporting configuration files.
The example assessment identifies potential weaknesses in delegation controls, budget propagation, tool permissions and human approval mechanisms.
Overall assessment: Improvements required before production deployment.
Important: The system, findings, file paths and scores below are fictional examples demonstrating how an audit report is structured.
2. Evidence Register
| Claim | Example source | Evidence type | Status |
|---|---|---|---|
| Tool-call authorisation | src/gateway/auth.py:42 | Code | Verified in scenario |
| Subagent budget limits | config/agents.yaml | Design | Contradicted in scenario |
| Human approval gate | src/orchestrator.py | Code | Not found in scenario |
| Tool restrictions | prompts/system_msg.txt | Prompt | Advisory only in scenario |
Evidence statuses describe the fictional scenario, not independently verified files.
3. Domain Scorecard
| Audit domain | Example score | Assessment |
|---|---|---|
| Tool permissions | 3/5 | Server-side controls exist but lack granular scopes |
| Delegation safety | 2/5 | Subagent identity propagation is incomplete |
| Reliability | 2/5 | Retry behaviour requires improvement |
| Observability | 4/5 | Logging is comprehensive but requires review |
| Remaining audit domains | Unknown | Insufficient example evidence |
Scoring principle: Missing evidence is reported as Unknown rather than assumed to pass or fail.
4. Key Findings
Finding AF-001 — Missing Budget Propagation
Severity: High
Example evidence: config/agents.yaml defines a global execution budget, but the example orchestrator does not consistently propagate that budget to delegated agents.
Risk: Child agents could exceed intended usage limits or generate unexpected inference costs.
Recommended remediation:
- Enforce budget ceilings at the orchestration layer.
- Propagate remaining budgets to child tasks.
- Validate spending limits server-side.
- Add tests for exhausted budgets and failed delegation.
Finding AF-002 — Advisory-Only Tool Restrictions
Severity: Critical
Example evidence: prompts/system_msg.txt instructs an agent not to invoke a destructive tool, but the fictional gateway lacks an independently enforced permission rule.
Risk: An agent could attempt an unauthorised operation if the restriction exists only in prompt instructions.
Recommended remediation:
- Implement server-side tool authorisation.
- Apply least-privilege access policies.
- Validate agent identity on every tool request.
- Test that prohibited actions are rejected.
Finding AF-003 — Missing Human Approval Gate
Severity: High
Example evidence: The fictional orchestration workflow does not demonstrate a mandatory approval mechanism for actions involving production resources.
Risk: Sensitive operations could proceed without appropriate human oversight.
Recommended remediation:
- Introduce explicit approval gates.
- Record the approving identity and decision.
- Reject operations when approval is missing or expired.
- Test approval bypass and cancellation scenarios.
5. Architecture Visualisation
AgentFlow Auditor can represent the examined architecture using an evidence-aware system map.
Example system structure:
Human → CoordinatorAgent → Tool Gateway → ToolExecutor → Data Store
The optional interactive viewer supports
- A 2D architecture view.
- A perspective-style system map.
- Inspection of nodes and connections.
- Evidence and verification-status indicators.
- Identification of unknown or contradictory relationships.
Map contents depend on the architecture evidence provided. Optional local generation requires a compatible environment and user authorisation.
6. Prioritised Remediation Plan
| Milestone | Action | Verification |
|---|---|---|
| M1 — Access control | Enforce server-side tool permissions | Unauthorised tool requests are rejected |
| M2 — Delegation | Propagate identity and budget limits | Child agents cannot exceed delegated authority |
| M3 — Human oversight | Implement approval gates | Sensitive actions fail without valid approval |
| M4 — Reliability | Improve retry and cancellation handling | Negative-path tests pass |
| M5 — Release verification | Review CI and release controls | Required build and approval gates are enforced |
7. Final Assessment
This fictional example illustrates how AgentFlow Auditor produces:
- Evidence-linked findings and an audit register.
- Risk and severity classifications.
- An eight-domain assessment framework.
- Explicit Unknown statuses for missing evidence.
- Prioritised recommendations and verification criteria.
- Optional architecture visualisation.
- Milestone-based remediation plans.
Limitations: This demonstration does not represent an actual security assessment. ChatGPT-assisted testing and local automated tests do not replace end-to-end testing in every supported AI-agent environment, independent security verification or production certification.
Connects securely to your tools. The creator never sees your data.
About this skill
AI agent architectures often fail due to brittle tool delegation, missing human-in-the-loop gates, and poor observability. This skill conducts an evidence-based audit of your agentic workflows, identifying security gaps and reliability risks before they hit production.
What it does
- Map execution paths by identifying entrypoints, agent identities, orchestration flows, and tool gateway enforcement points.
- Verify handover claims using a claim-to-source matrix that tracks design intent versus actual code implementation.
- Audit eight domains including tool permissions, delegation safety, failure handling, and cost controls using a standardized rubric.
- Classify findings with specific exploit scenarios, impact analysis, and reproducible verification steps.
- Generate architecture maps that visualize nodes and edges while distinguishing between verified and unknown relationships.
How it works
- Submit inputs such as repository files, design notes, agent traces, or architecture diagrams for a read-only review.
- Review the evidence register to see which claims are supported by code and which remain unverified hypotheses.
- Analyze the scorecard to view scores across assessed domains without fabricated grades for missing data.
- Execute remediation based on a prioritized plan focusing on credential leaks, unauthorized paths, and reliability.
Frameworks & tools
Works with any agentic framework including LangChain, AutoGen, and CrewAI, or custom Claude Code subagent implementations. Uses structured rubrics, threat models, and optional Python scripts for deterministic scoring and HTML visualization.
Why this beats prompting it yourself
Generic prompts often miss the distinction between a prompt-based instruction and a server-side enforcement boundary. This skill enforces a strict evidence-first methodology that labels unverified claims as "unknown" rather than assuming safety, preventing false confidence in your agent's security.
Use cases
- Audit orchestrator files to find missing guardrails in complex multi-agent delegation flows.
- Review PRs for new agent capabilities to ensure tool permissions follow the principle of least privilege.
- Assess analytics agents to verify data lineage, metric definitions, and reproducibility of generated reports.
- Analyze failure handling by inspecting retry idempotency, timeouts, and budget propagation across sessions.
Known limitations
This is a read-only assessment and does not perform active penetration testing or execute repository scripts. Findings are based on provided evidence; lack of evidence is reported as an assurance gap.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
AgentFlow Auditor performs evidence-based, read-only assessments by default. It does not perform penetration testing, execute the target application, or guarantee security or compliance. Results depend on the quality and completeness of supplied evidence. Optional local scoring and architecture visualisation require a compatible environment and explicit permission to run.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 19 days ago