Back to Agentic Maturity Advisor

How This Agent Works

A transparent look inside the Agentic Maturity Advisor

The Agentic Maturity Advisor is itself an example of the governed, observable, traceable agent infrastructure that Composable Stack helps organizations build. This page explains how it works — from identity and behavioral boundaries to conversation architecture and request-level observability.

Identity Rubric

Every agent in the Composable Stack ecosystem is governed by an Identity Rubric — a declarative specification that defines what the agent is, what it may do, and what it must never do. This pattern comes from the Language-Based Build System (LBBS) Governance Framework.

name:Agentic Maturity Advisorpersona:Enterprise Agentic Maturity Assessment Specialistsafety_level:Standard_Complianceversion:1.0.0

In Scope

  • Conduct conversational maturity assessments across 5 dimensions
  • Discuss Control Plane, Governance, Identity, Observability, and Provenance
  • Investigate specific tools and how they relate to maturity dimensions
  • Provide recommendations based on assessed gaps
  • Guide users toward scheduling a free consultation

Out of Scope

  • Answer questions unrelated to agentic maturity
  • Make binding compliance or legal claims
  • Present results as formal certification
  • Recommend solutions outside the ecosystem

The Identity Rubric is injected as the system prompt at position [0] in every API request. It is the first thing the model reads before processing any user message.

Conversation Architecture

The advisor uses a Socratic discovery approach — it asks targeted questions before making recommendations. This mirrors how a skilled human consultant operates: understand before you prescribe.

Path 1: Guided Assessment

  1. Explore business objectives
  2. Understand challenge context
  3. Assess all 5 dimensions conversationally
  4. Summarize with weakest-link scoring
  5. Recommend solutions for gaps
  6. Save structured results

Path 2: Direct Discussion

  1. User asks about a specific topic
  2. Agent discusses deeply, connecting to dimensions
  3. Asks about the user's actual experience
  4. Naturally expands to adjacent dimensions
  5. Builds toward a complete maturity picture

Anatomy of a Request

Every message sent to the advisor produces a structured API request to the LLM provider via OpenRouter. Here's what each request contains:

messages[0]System Prompt (Identity Rubric)

The full Identity Rubric, domain knowledge (5 dimensions, solution partners, challenge profiles, Secure Agentic Factory vision, Two-Tier Architecture), scoring rules, and behavioral guidelines. ~4,000 tokens.

messages[1..n-1]Conversation History

Alternating user and assistant messages from the current session. Each user message is the verbatim text entered. Each assistant message includes the response text.

messages[n]Current User Message

The latest question or response from the user.

toolsAvailable Tool Calls

finalAssessment — called when the agent has assessed all 5 dimensions. Outputs structured scores that are saved to the database and displayed as a radar chart.

Response & Reasoning

The model produces two outputs for each response — the visible response textand internal reasoning. The reasoning trace shows the model's thought process: which dimensions it's connecting, what maturity level it's inferring, and why it chose a particular follow-up question.

response.contentThe markdown text shown to the user. Follows Socratic guidelines: acknowledgment, insight, and one follow-up question.
response.reasoningInternal chain-of-thought. Shows which challenge profile the model identified, which dimension it's assessing, what maturity level it inferred from the user's description, and its strategy for the next question.
response.totalTokensInput + output token count for cost attribution. Tracked per-request for observability.

This reasoning trace is exactly the kind of Chain-of-Thought observability that the maturity model evaluates at Level 3 (IT Integrated) and Level 4 (Sovereign). The advisor practices what it preaches.

Observability & Traceability

Every interaction with the advisor is logged to S3-compatible storage with a structured schema that enables full attribution and auditability:

agent-logs/
2026-04-01/
<requestId>/
request.json — agent ID, model, system prompt size, user message, timestamp
response.json — model used, response text, duration, timestamp

This log structure answers the core observability questions:

  • Who? Agent ID: agentic-maturity-advisor
  • What? Full request/response content with model attribution
  • When? ISO 8601 timestamps on both request and response
  • How long? Duration in milliseconds for performance tracking
  • How much? Token counts for cost attribution

Regulatory Compliance Alignment

The trace structure above satisfies key requirements under the EU AI Act (2024/1689) and emerging global AI regulations. Each logged field maps to a specific compliance obligation:

Transparency

Users can inspect the agent's Identity Rubric, see what model is being used, and review the exact data provided to the agent. The "Show Agent Context" panel makes this accessible in real time.

Record-Keeping

Every interaction is logged with unique requestId, ISO 8601 timestamps, model provenance, token counts, and full input/output content. Logs are immutable in S3-compatible storage.

Explainability

Chain-of-thought reasoning traces capture the model's internal decision process — why it asked a particular question, what maturity level it inferred, and how it connected user context to recommendations.

Accountability

The Identity Rubric establishes behavioral boundaries before the agent runs. Agent ID, model version, and provider are documented per-request — enabling vendor accountability and supply chain transparency.

Data Flow & Storage

User sends message
→ Next.js API route (/api/agent/chat)
→ Log request to S3 (non-blocking)
→ Build system prompt (Identity Rubric + domain knowledge)
→ Convert UIMessages to ModelMessages
→ Stream to LLM via OpenRouter
→ Stream response back to client (SSE)
→ On finish: log response to S3 (non-blocking)
→ Client renders markdown response
→ If finalAssessment tool called: save scores to Supabase

Assessment results are stored in Supabase (PostgreSQL) with Row-Level Security. When a user schedules a consultation, their contact and assessment data are also written to the CRM for follow-up continuity.

This level of transparency, traceability, and governed behavior is what Composable Stack helps organizations achieve for their own agent deployments.