MCP and the elegant agentic loop

The elegant agentic loop describes how WCGW Safety Suite moves from graph construction to audit. A human engineer builds a traceability graph of work items, concepts, and their relations. AI agents read that graph, scan for gaps and inconsistencies, and return suggestions with ready-to-apply payloads. You review each suggestion and sign the ones that hold up; they become revisions. Accepted revisions are bundled into baselines that can be compared and audited.

The same loop is open to any compatible AI client through the WCGW MCP server. Every step below has a matching MCP tool, so an agent can run the loop with you from the client you already use.

Suggestions are attributed and revisions are signed, so an auditor can trace any concept back through the audit trail to the revision that introduced it and the requirement that justified it.

flowchart TB
  classDef human fill:#4a90d9,stroke:#2c5f8a,color:#fff
  classDef agent fill:#9b59b6,stroke:#6c3483,color:#fff
  classDef graphStyle fill:#27ae60,stroke:#1e8449,color:#fff
  classDef core fill:#16a085,stroke:#0e6655,color:#fff
  classDef complete fill:#e67e22,stroke:#d35400,color:#fff
  classDef artifact fill:#7f8c8d,stroke:#5d6d7e,color:#fff

  subgraph H["Human Layer"]
    E[Engineer]:::human
    A[Auditor]:::human
  end

  subgraph GL["Graph Layer"]
    W[Work items]:::graphStyle
    C[Concepts & relations]:::graphStyle
    R[Concept revisions]:::graphStyle
    T[Templates]:::graphStyle
    G[Traceability graph]:::core
  end

  subgraph AG["AI Agent layer"]
    EA[Evaluation agent]:::agent
    QA[Query agent]:::agent
    MC[MCP client]:::agent
  end

  subgraph AL["Artifact Layer"]
    SUG[Suggestions]:::artifact
    BS[Baseline snapshot]:::artifact
    AT[Audit trail]:::artifact
    M[Completeness metrics]:::complete
  end

  E -->|builds| W
  W -->|links| C
  C -->|versions| R
  T -->|seeds| C

  W -->|constitutes| G
  C -->|constitutes| G
  R -->|constitutes| G

  G -->|scans| EA
  G -->|reads| QA
  QA -->|enriches| G
  E -->|works with| MC
  MC -->|calls tools on| G

  EA -->|proposes| SUG
  SUG -->|modifies| R
  SUG -->|modifies| C
  E -->|acts on| SUG

  G -->|measures| M

  R -->|snapshots| BS
  BS -->|records| AT

  A -->|verifies| AT
  A -->|compares| BS
  A -->|reads| M
  A -->|queries| G

The layers

How the loop runs

  1. You build work items and link concepts into a traceability graph.
  2. Templates seed concepts; revisions capture each change and constitute the graph.
  3. Completeness metrics measure how close a work item is to audit-ready.
  4. The evaluation agent proposes suggestions; the query agent reads and enriches the graph.
  5. You review and sign suggestions that hold up; they become revisions.
  6. Accepted revisions are captured in a baseline snapshot, which records the audit trail.
  7. The auditor verifies the audit trail, compares baselines, reads completeness metrics, and queries the graph.

You use the system by building a traceability graph, running queries to find gaps, and reviewing the suggestions the agent proposes. Sign the ones that hold up, and bundle approved revisions into baselines. The auditing body verifies the audit trail, compares baseline snapshots, reads completeness metrics, and queries the graph to confirm every change was reviewed and attributed.

Running the loop over MCP

The Model Context Protocol (MCP) lets an AI client call WCGW Safety Suite directly. The client signs in with OAuth as you, so every tool call acts with your permissions and is attributed to your account. The agent never sees projects you are not a member of. Writes need an active license, and AI-assisted tools consume an LLM call.

Each step of the loop maps to MCP tools:

Loop step MCP tools
Find your project and work list_my_projects, list_work_items, get_work_item
Build the graph create_work_item, list_work_item_templates, create_concept, create_relation, add_concept_to_work_item
Measure completeness get_graph_completeness, get_refinement_report
Query for gaps get_graph_query_categories, plan_graph_query_sequence, run_graph_query
Evaluate and get suggestions evaluate_work_item, get_evaluation_result, list_evaluation_results
Review and sign act_on_suggestion, apply_evaluation_suggestions, submit_concept_revision, list_concept_revisions
Baseline for audit create_project_baseline, list_project_baselines, get_project_baseline

The published safety engineering skills, such as Walk the Chain and STPA, are also exposed as MCP tools. They let an agent run a structured analysis over the same graph.

The agent's role in the loop stays the same: it reads, queries, and proposes. Suggestions only become revisions when you act on them, and the audit trail records who signed each one. An agent connected over MCP does not certify compliance. It does the legwork so that you have more time for engineering judgement.

Connecting a client