Trial of the Agentic Codex: The Grand Capstone

The Hall of Mastery stands silent. The six seals of the Codex glow on the walls — one for each domain, one for each truth the candidate has learned on the journey from initiate to master. The Codex Master who placed the seals speaks: “Show me that you did not only read the Codex. Show me that you understand why it was written.”

🗺️ The Arc Complete

graph TD
    D1["Domain 1: Agentic SDLC"]
    D2["Domain 2: Tools & Permissions"]
    D3["Domain 3: Memory & Context"]
    D4["Domain 4: Evaluation"]
    D5["Domain 5: Multi-Agent"]
    D6["Domain 6: Governance"]
    CAP["🏆 CAPSTONE: Agentic Codex Master"]
    
    D1 --> D2 --> D3 --> D4 --> D5 --> D6 --> CAP
    
    style CAP fill:#FFD700,stroke:#B8860B,stroke-width:4px,color:#000
    style D1 fill:#4CAF50,stroke:#2E7D32,color:#fff
    style D2 fill:#4CAF50,stroke:#2E7D32,color:#fff
    style D3 fill:#4CAF50,stroke:#2E7D32,color:#fff
    style D4 fill:#4CAF50,stroke:#2E7D32,color:#fff
    style D5 fill:#4CAF50,stroke:#2E7D32,color:#fff
    style D6 fill:#4CAF50,stroke:#2E7D32,color:#fff

🎯 Capstone Objectives

• Seal 1 (Domain 1 — 18%): Implement agent-in-SDLC and define boundaries
• Seal 2 (Domain 2 — 18%): Configure tools, permissions, MCP, environment integration
• Seal 3 (Domain 3 — 19%): Implement memory strategy and context continuity
• Seal 4 (Domain 4 — 19%): Evaluate agent performance and iterate on instructions
• Seal 5 (Domain 5 — 17%): Build and manage a multi-agent system
• Seal 6 (Domain 6 — 9%): Implement responsible autonomy, guardrails, and HITL

🏛️ The Grand Trial Scenario

The Scribe presents the scenario:

You are the lead AI engineer at a software team that has decided to adopt agentic AI development using GitHub. You have been given an empty repository, a GitHub account with Copilot, Actions, and Environments, and a mandate: build a working agentic SDLC in the next 6 hours and demonstrate competency in every domain.

The following 6 chapters map directly to the GH-600 exam domains.

⚔️ Seal 1: The Agentic SDLC (Domain 1 — 18%)

Related quests: Q1 (SDLC Integration), Q2 (Plan vs Action), Q3 (Observability)

Challenge 1.1: Describe where agents live in your SDLC

Task: Document where in your software development lifecycle agents operate. Produce a diagram.

<!-- work/gh-600/capstone/sdlc-diagram.md -->
# Our Agentic SDLC

## Agent Integration Points

| Phase | Agent Role | Trigger | Human Touchpoint |
|---|---|---|---|
| Planning | Requirements analysis | Issue created | Human approves specification |
| Implementation | Code writing | `agent-implement` label | Human reviews PR |
| Review | Code review comments | PR opened | Human accepts/rejects suggestions |
| Testing | Test execution | PR opened | Human reviews failures |
| Deployment | Deploy staging | PR merged to main | Human approves production |

## Architecture Diagram
[Include Mermaid diagram here]

Challenge 1.2: Demonstrate planning vs. action separation

Task: Show a GitHub Actions workflow that separates the plan step from the execute step with a mandatory break between them.

# .github/workflows/plan-then-execute.yml
name: Plan-Then-Execute (Sealed Capstone)

on:
  issues:
    types: [labeled]

jobs:
  plan:
    if: contains(github.event.label.name, 'agent-implement')
    runs-on: ubuntu-latest
    outputs:
      plan_approved: $
    steps:
      - name: Generate plan
        id: plan
        run: |
          echo "Generating implementation plan..."
          # Agent generates plan here — does NOT execute
          echo "approved=pending" >> "$GITHUB_OUTPUT"

      - name: Post plan for approval
        uses: actions/github-script@v7
        with:
          script: |
            await github.rest.issues.createComment({
              owner: context.repo.owner,
              repo: context.repo.repo,
              issue_number: context.payload.issue.number,
              body: '**Agent Plan Generated** ✅\n\nReact with 👍 to approve execution, or 👎 to reject.\n\n_No changes have been made yet._'
            });

  execute:
    needs: plan
    runs-on: ubuntu-latest
    environment: agent-approval    # Human must approve in GitHub UI before this runs
    steps:
      - name: Execute approved plan
        run: echo "Executing plan after human approval..."

Challenge 1.3: Configure an observability workflow

Task: Ensure every agent run emits structured logs that can be queried.

Refer to Q3 (Observability & Control) for the full pattern.

⚔️ Seal 2: Tools, Permissions, and Environment (Domain 2 — 18%)

Related quests: Q4 (Tool Selection), Q5 (MCP), Q6 (Dev Env), Q7 (Safe Execution)

Challenge 2.1: Configure a scoped GitHub token

Task: Set the minimum permissions needed for an implementation agent.

# In your workflow:
permissions:
  contents: write        # Write files to repository
  pull-requests: write   # Create and update PRs
  issues: write          # Comment on issues
  # Explicitly deny everything else:
  # actions: none (default)
  # security-events: none (default)

Challenge 2.2: Configure an MCP server

Task: Add a GitHub MCP server to Copilot and demonstrate its use.

// .vscode/mcp.json (workspace-level MCP configuration)
{
  "servers": {
    "github": {
      "command": "npx",
      "args": ["@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${input:github-token}"
      }
    }
  }
}

Challenge 2.3: Document error handling and escalation

Task: Add a failure escalation workflow that creates an issue when an agent fails.

Refer to Q7 (Safe Execution & Error Handling) for the full pattern.

⚔️ Seal 3: Memory and Context (Domain 3 — 19%)

Related quests: Q8 (Memory Strategies), Q9 (State Persistence), Q10 (Cross-tool Continuity)

Challenge 3.1: Implement all three memory tiers

Task: Create artifacts demonstrating ephemeral, session, and persistent memory.

## Memory Tier Implementation Checklist

- [ ] **Ephemeral**: Variables in workflow `env:` block used within a job
- [ ] **Session**: Artifact uploaded in step A, downloaded in step B
- [ ] **Persistent**: Repository file updated by agent, committed and pushed

Challenge 3.2: Implement drift detection

Task: Produce a drift check that compares current agent state to expected state.

Refer to Q9 (State Persistence & Drift) for detect_drift.py.

Challenge 3.3: Implement cross-surface context handoff

Task: Show a context-handoff.json passed between an issue → PR → branch context.

Refer to Q10 (State Continuity Cross-Tools) for the schema and inject_cross_surface_context.py.

⚔️ Seal 4: Evaluation and Performance (Domain 4 — 19%)

Related quests: Q11 (Success Criteria), Q12 (Root Cause Analysis), Q13 (Behavior Tuning)

Challenge 4.1: Define machine-verifiable acceptance criteria

Task: Write 3 acceptance criteria for an agent task that can be verified programmatically.

// work/gh-600/capstone/acceptance-criteria.json
{
  "task": "Implement authentication feature",
  "criteria": [
    {
      "id": "AC-01",
      "description": "Unit tests pass",
      "signal": "ci-pass",
      "check_command": "gh run list --workflow=test.yml --branch=feature/auth --status=success --limit=1"
    },
    {
      "id": "AC-02",
      "description": "No new security vulnerabilities",
      "signal": "security-scan-pass",
      "check_command": "gh api /repos/{owner}/{repo}/code-scanning/alerts?state=open | jq 'length == 0'"
    },
    {
      "id": "AC-03",
      "description": "Code review approved",
      "signal": "pr-approved",
      "check_command": "gh pr view {pr_number} --json reviewDecision -q '.reviewDecision == \"APPROVED\"'"
    }
  ]
}

Challenge 4.2: Perform an RCA on a failed run

Task: Take a failed workflow run and produce a 5-Why RCA document.

Refer to Q12 (Failure Root Cause Analysis) for the full RCA template.

Challenge 4.3: Iterate on agent instructions

Task: Make one instruction change, measure the before/after difference.

Refer to Q13 (Behavior Tuning) for the instruction changelog template.

⚔️ Seal 5: Multi-Agent Systems (Domain 5 — 17%)

Related quests: Q14 (Orchestration), Q15 (Observability), Q16 (Recovery), Q17 (Lifecycle)

Challenge 5.1: Design a 3-agent orchestration workflow

Task: Create a fan-out or chain orchestration with 3 sub-agents.

Refer to Q14 (Multi-Agent Orchestration Patterns) for the fan-out and chain patterns.

Challenge 5.2: Add tracing to the multi-agent run

Task: Each sub-agent emits a trace entry with a shared correlation ID.

Refer to Q15 (Multi-Agent Observability) for trace_writer.py.

Challenge 5.3: Add failure recovery to the orchestrator

Task: Orchestrator continues after one sub-agent fails and produces a recovery plan.

Refer to Q16 (Multi-Agent Failure Recovery) for recovery_coordinator.py.

Challenge 5.4: Register all agents in the agent registry

Task: Publish _data/agents.yml with all 3 agents registered.

Refer to Q17 (Multi-Agent Lifecycle Management) for the registry schema.

⚔️ Seal 6: Responsible Agentic AI (Domain 6 — 9%)

Related quests: Q18 (Autonomy Levels), Q19 (Guardrails & HITL)

Challenge 6.1: Produce your autonomy matrix

Task: Complete _data/autonomy-matrix.yml with 5 task types at appropriate levels.

Refer to Q18 (Autonomy Levels Matrix) for the matrix schema.

Challenge 6.2: Implement 3 guardrails

Task: CODEOWNERS file-scope boundary, approval gate environment, forbidden actions list.

Refer to Q19 (Guardrails & Human-in-the-Loop) for each guardrail.

📋 Domain Coverage Rubric (GH-600 Exam Alignment)

🪞 The Grand Reflection

After completing all 6 seals, publish your reflection:

<!-- work/gh-600/capstone/grand-reflection.md -->
# Grand Reflection: Agentic Codex Trial

## What I Built
[Summary of the agentic system you designed and implemented]

## Most Challenging Domain
[Which domain was hardest and why?]

## Key Architectural Decision
[The most important decision you made and why]

## What I Would Do Differently
[Honest reflection on what could be improved]

## Exam Readiness Self-Assessment
[Domain-by-domain confidence rating 1-5]

## Resources I Would Review Again
[Links back to quests or docs that were most valuable]

✅ Capstone Validation

# Validate all 20 quests in the arc
python3 test/quest-validator/quest_validator.py -d pages/_quests/

# Check all 6 seals are present
python3 work/gh-600/scripts/validate_capstone.py \
  --registry _data/agents.yml \
  --matrix _data/autonomy-matrix.yml \
  --reflection work/gh-600/capstone/grand-reflection.md

# Build site
docker-compose exec jekyll bundle exec jekyll build

🏆 The Agentic Codex — Complete Arc Links

🏆 Capstone Rewards

The Codex Master speaks: “The seals are broken. The knowledge is yours. Go now and build systems worthy of the Codex.”