GH-600 in the Wild: Implemented in IT-Journey

The Agentic Codex is not a thought experiment — this repository runs on the disciplines the exam tests. An AI fleet drafts content, walks quests as a learner, repairs what it witnessed breaking, triages issues, and audits itself — every lane behind kill switches, deterministic gates, and a human merge button. This note maps each GH-600 domain to the live artifact that implements it, so you can read real source instead of imagining it.

How to use this page: study a chapter, then open the artifact and see the same pattern holding production weight. Each path links to the source on GitHub; each row names the chapter that teaches the concept.

The control plane in one workflow

.github/workflows/agent-plan-then-act.yml is the campaign’s reference pipeline, live: a plan job that only plans, an execute job held Waiting behind the agent-approval Environment, the plan crossing the boundary as an artifact, a drift-guard verify before acting, a correlation ID threading every step, and an audit artifact recording instruction → action → outcome. It is OFF by default behind AGENT_DEMO_ENABLED, exactly like every other agent lane here.

Domain 1 — Agentic AI in the SDLC (18%)

Sub-skill	Implemented by	What it proves
Bounded agents with defined entry/exit	_data/agents/registry.yml + .claude/agents/	Nine named agents, each with ONE role, a lane (read-only vs content-write), and a kill switch
Plan vs action separation	agent-plan-then-act.yml; walk→fix split in quest-perfection.yml	The quest loop’s walker only witnesses (read-only); a separate fixer acts — planning and acting are different jobs, even different workflows
Observability & control	Workflow step summaries + committed .quests/ledger.json + .quests/DASHBOARD.md	Every autonomous run leaves a trace a human can audit without re-running

Taught by Chapter I — Initiation Rites.

Domain 2 — Tool Use & Environment (18%)

Sub-skill	Implemented by	What it proves
Tool selection & least privilege	permissions: blocks in every workflow (default contents: read); scripts/ai/run.sh --tools allow-list	Grants are per-job and by omission; the runner passes agents only the tools the task names
MCP server configuration	.vscode/mcp.json (editor, promptString token) + scripts/ai/mcp/github-readonly.json (runner, env token)	Declared servers, secrets never committed, allow-list at the call site
Environment integration	AGENTS.md + .github/copilot-instructions.md + _data/ai.yml	The realm’s law, written for the agent; ONE config file for every model call
Safe execution paths	Claude-Code→API fallback in run.sh; timeout-minutes + concurrency on every lane; needs-human escalation	Failures retry once (fallback), then surface loudly — never a silent green

Taught by Chapter II — Forging the Arsenal.

Domain 3 — Memory, State & Execution (19%)

Sub-skill	Implemented by	What it proves
Three memory tiers	Job outputs (ephemeral) · walk-plan/evidence artifacts between walk and fix jobs (session) · committed .quests/ledger.json + .cms/ index (persistent)	State lives in the tier that matches how long it must survive
Durable progress, resumability	The quest-perfection ledger: each slice’s status survives across daily runs; stuck/needs_human slices are skipped, not repeated	The next run reads the register and continues instead of repeating
Context-drift detection	scripts/ai/drift-guard.sh, wired between plan and act in the reference pipeline	Snapshot at plan time, verify before acting, exit 78 aborts on a moved world

Taught by Chapter III — Vaults of Recollection.

Domain 4 — Evaluation & Tuning (19%)

Sub-skill	Implemented by	What it proves
Machine-verifiable success criteria	scripts/quest/quest_audit.py (tier-1 score ≥ 70%) · scripts/ci/brand_lint.py · scripts/ci/classify_changes.py	“Done” is a function of signals, not an opinion — drift blocks the merge
The model never grades itself	The deterministic keep/revert gate in quest-fix.yml	An edit is kept only if the deterministic signal improved — never on the model’s own grade
Failure RCA	docs/agents/rca-template.md + gh run download forensics convention	Classify by layer, 5-Why to a root cause in our system, then ship the prevention
Instructions as code	docs/agents/instructions-changelog.md	Every behavior change carries its reason and its measured outcome

Taught by Chapter IV — The Oracle Rubric.

Domain 5 — Multi-Agent Coordination (17%)

Sub-skill	Implemented by	What it proves
Orchestration patterns	quest-perfection.yml chains walk → fix per slice; the fleet’s daily lanes run in parallel isolation (one PR each, non-overlapping scopes)	Chain for dependent work, fan-out for independent work, reconciliation at the join (auto-merge lanes)
Distributed tracing	Correlation ID threaded through jobs, summaries, and artifact names in agent-plan-then-act.yml; slice ids + ledger entries across walk/fix	One identifier reassembles a multi-job run into a single narrative
Failure recovery	continue-on-error/skip semantics in the perfection loop; ledger marks slices stuck/needs_human (escalate) instead of blind retry	Stalled vs conflicting handled differently; escalation is a first-class outcome
Lifecycle management	_data/agents/registry.yml + weekly agent-audit.yml	Provision by row, deprecate by status flip, audited quarterly (review_date)

Taught by Chapter V — The Council of Many.

Domain 6 — Guardrails & Accountability (9%)

Sub-skill	Implemented by	What it proves
Risk-based autonomy levels	_data/agents/autonomy-matrix.yml	Every recurring action classified L0–L4 by reversibility, blast radius, predictability — with the guardrails that justify the level
File-scope boundary	.github/CODEOWNERS	The workflow tree and quest framework demand a named human reviewer
Human approval gate	The agent-approval Environment in the reference pipeline; every AI lane’s *_ENABLED kill switch + auth secret (both required, OFF by default)	Autonomy is opt-in twice, and the irreversible half waits for a human
Forbidden-actions pact	The Warden Pact in AGENTS.md	A constraint that holds regardless of instruction — decline, comment, label, stop
Audit trail	Committed ledger + Actions run logs + PR history + the audit artifact in the reference pipeline	Instruction, action, and outcome reconstructable from the repo’s own history

Taught by Chapter VI — The Warden Pact.

Try it yourself

1. Read a chapter, then its artifact above — the pattern is identical, at scale.
2. Run the reference pipeline in your fork: create the agent-approval Environment with yourself as required reviewer, set AGENT_DEMO_ENABLED=true, dispatch 🜂 Agent Plan-then-Act, and watch the execute job wait for your seal.
3. Face the Grand Capstone — this repository is a worked answer key, but the trial wants your build.

Related

• Epic Quest: The Agentic Codex — the campaign hub
• GH-600 Study Hub — domains, weights, and the quest line
• Epic Quest: The Self-Operating Website — the campaign that builds a fleet like this one