Ai Agent Architecture Audit

Audit AI-agent architecture

Assess whether probabilistic model behavior is safely bounded by deterministic product and operational controls.

Load the contract

Read these files completely:

• ../../resources/references/review-contract.md
• ../../resources/references/knowledge-contract.md
• ../../resources/references/ai-agent-rules.md
• ../../resources/rules/ai-agent-core.yaml
• ../../resources/knowledge/manifest.yaml

Load the project's profile, constraints, and critical flows. Use project-architecture-audit as well when the request covers the whole product rather than only AI-specific boundaries.

Workflow

1. Require a valid project Profile before the specialized audit. When

.architecture/profile.yaml is missing and persistence is allowed, run architecture_tool.py prepare-project-audit --repo , then validate the project. When the user explicitly requests read-only work, do not create governance; report the missing Profile and stop with a project-architecture-audit bootstrap handoff.

1. Set one stable `` for the audit and preserve every generated input

under .architecture/reviews/inputs/; never reuse or overwrite a prior Review's evidence chain. Inspect current facts, build a current Profile from the declared Profile, then select Knowledge:

``bash python3 ../../resources/scripts/architecture_tool.py inspect-repository \ --repo \ --output /.architecture/reviews/inputs/-repository-facts.yaml python3 ../../resources/scripts/architecture_tool.py build-profile \ --facts /.architecture/reviews/inputs/-repository-facts.yaml \ --declared /.architecture/profile.yaml \ --output /.architecture/reviews/inputs/-profile.yaml python3 ../../resources/scripts/architecture_tool.py select-knowledge \ --facts /.architecture/reviews/inputs/-repository-facts.yaml \ --profile /.architecture/reviews/inputs/-profile.yaml \ --task "" \ --skill ai-agent-architecture-audit \ --output /.architecture/reviews/inputs/-knowledge-selection.yaml \ --context-output /.architecture/reviews/inputs/-knowledge-context.yaml python3 ../../resources/scripts/architecture_tool.py validate-knowledge-context \ /.architecture/reviews/inputs/-knowledge-context.yaml \ --selection /.architecture/reviews/inputs/-knowledge-selection.yaml \ --facts /.architecture/reviews/inputs/-repository-facts.yaml \ --profile /.architecture/reviews/inputs/-profile.yaml ``

Read the compact context projection only after validation succeeds; reserve the full exclusion ledger for scripts, Reviews, and Gates. Do not load every selected Markdown entry by default. Open a complete selected entry only after verifying its recorded hash and only when a candidate-driving claim, ambiguity, volatile fact, or explicit trade-off depends on its source detail. Do not load unrelated packs.

1. Draw the control path from user intent through orchestration, model calls, retrieval, tools, persisted state, human approval, and side effects.
2. Separate deterministic services and policy enforcement from model judgment.
3. Identify trust boundaries for user content, retrieved content, prompts, tools, credentials, memory, and model providers.
4. Inventory every context source and record its owner, purpose, necessity, authority, scope, freshness, sensitivity, transformation, retention, and disposal. Require an explicit reason for every included source and evidence for every omitted source that could affect a critical flow.
5. Trace context and memory lifecycles: creation, scoping, provenance, retention, mutation, retrieval, deletion, and recovery. Inspect budget, truncation, summarization, ranking, and fallback behavior; require preservation or explicit rejection of authority, provenance, and recency.
6. Compare context ordering across retries, tasks, tenants, and releases. Inspect cache keys, cache boundaries, invalidation, and reuse so ordering changes or volatile data cannot silently cross an authorization or provenance boundary.
7. Classify stable policy, contracts, and instructions separately from volatile user input, retrieval results, task state, and provider responses. Prevent volatile or sensitive material from entering stable context or reusable caches.
8. Minimize sensitive and personal data independently at prompt, retrieval, memory, and trace boundaries. Inspect field-level allowlists, redaction, purpose limitation, scope, retention, deletion, and reference-versus-raw-content choices.
9. Trace long-running task state, idempotency, retries, cancellation, checkpoints, resumption, and duplicate side effects.
10. Inspect model routing, timeouts, fallbacks, degraded modes, cost budgets, latency budgets, and failure visibility.
11. Bind behavior evidence to exact model/runtime, prompt, tool policy and schema, retriever/index/ranking configuration, context treatment, evaluation data, environment, timestamp, and hashes where available.
12. For any new or upgraded agent runtime, protocol, model, or framework, compare adopt, retain, and reject decisions against the current baseline using critical-flow evidence for quality, compatibility, security, operations, cost, ownership, rollout, and rollback.
13. Inspect evidence capture, source attribution, evaluation coverage, production feedback, and human confirmation boundaries.
14. Assess every applicable AI rule and record explicit not_applicable or not_assessed states.

For each finding or strength, bind the assessment to at least one concrete affected critical flow and prove the complete trigger-to-impact or control-to-outcome path. State findings as architecture invariants and affected capabilities; record technology names only as versioned evidence. A prompt containing untrusted text is not by itself a prompt-injection vulnerability; show how it can cross a policy or tool boundary.

Verification handoff and output

Apply the candidate evidence requirements in review-contract.md. Leave every finding at verification.status: candidate. Do not turn this candidate handoff into a verified conclusion; preserve the candidate/verification boundary and hand off confirmation to the independent verifier.

Write persistent artifacts under .architecture/reviews/ using kind ai-agent:

• -ai-agent-candidates.yaml;

Start machine-readable output from ../../resources/templates/review.yaml and set review.kind to ai-agent.

Use Review schema 1.2. Bind the exact per-run repository-facts, Profile, and AI knowledge-selection paths and hashes, preserve fact/inference boundaries, enumerate critical-flow coverage, and validate with:

python3 ../../resources/scripts/architecture_tool.py validate-review \
   --project 
python3 ../../resources/scripts/architecture_tool.py validate-coverage \
  --project  --review  --allow-candidates

Hand off architecture, candidate strengths and risks, critical-flow impact, coverage, counter-evidence, and limitations. Use $architecture-finding-verifier for confirmed conclusions and the final report. Do not prescribe fixes.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: qingye-lab
• Source: qingye-lab/hengmu
• License: MIT
• Homepage: https://qingye-lab.github.io/hengmu/

Install and usage instructions live in the source repository linked above.