Agentic Coding Hub: Workflows, Context Anchoring, and Autonomous Fleets

This cluster covers the practice of building software with AI coding agents (Antigravity agy, Claude Code, Cursor, Codex, Devin) — the architectural workflows, context engineering disciplines, and deterministic verification loops that separate rigorous agentic software engineering from brittle autocomplete.

The governing invariant of modern agentic engineering: frontier models have largely converged in raw code generation capabilities; sustainable leverage is determined entirely by the context architecture and verification harness surrounding the agent. Two engineers deploying the same frontier reasoning model achieve an order-of-magnitude difference in velocity and defect escape rates depending on how they anchor intent, assemble context, gate execution, and parallelize across worktrees.

flowchart TD
    subgraph Sovereign_Knowledge["1. Sovereign Knowledge Plane (Wikantik)"]
        WIKI[("Semantic Wiki Knowledge Base<br/>(living specs, ADRs, schemas)")]
        MCP_K["Knowledge MCP Server<br/>(/knowledge-mcp, 21 read tools)"]
        WIKI -->|assemble_bundle / get_page_for_agent| MCP_K
    end

    subgraph Agent_Harness["2. Agent Context & Execution Engine"]
        RULES["Root Rules & Headers<br/>(AGENTS.md / CLAUDE.md)"]
        SKILLS["Procedural Skills Library<br/>(.agents/skills/ / SKILL.md)"]
        AGENT{"Coding Agent Runtime<br/>(Antigravity agy / Claude Code)"}
        
        RULES --> AGENT
        SKILLS --> AGENT
        MCP_K -->|Authoritative Context Injection| AGENT
    end

    subgraph Implementation_Plane["3. Local Implementation & Gating"]
        WT["Isolated Git Worktree<br/>(Branch isolation)"]
        TESTS["Test-First Harness (TDD)<br/>(Unit, integration, mutation)"]
        LINT["Static Analysis & Linters<br/>(Checkstyle, PMD, SpotBugs)"]
        AGENT -->|Code Edits & Bash| WT
        WT --> TESTS
        TESTS --> LINT
    end

    subgraph Feedback_and_Close["4. Verification & Living Drift Closure"]
        MCP_A["Admin MCP Server<br/>(/wikantik-admin-mcp, 29 tools)"]
        LINT -->|Pass Gates| MERGE[PR Review / Clean Merge]
        LINT -->|Fail Gates| AGENT
        MERGE -->|Update Specs / mark_page_verified| MCP_A
        MCP_A --> WIKI
    end

1. The Three Eras of AI-Assisted Coding

Engineering organizations advance through three distinct architectural paradigms:

Dimension	Era 1: Vibe Coding (Ad-Hoc Prompting)	Era 2: In-Repo Spec-Driven Development	Era 3: Semantic Wiki-Anchored Engineering (Level 5)
Source of Truth	Ephemeral chat scrollback and human memory	Markdown files in repository (`specs/plan.md`)	Centralized semantic wiki with bidirectional MCP
Context Assembly	Manual copy-pasting of files and logs	Static `CLAUDE.md` / `AGENTS.md` instruction bloat	Token-budgeted semantic bundle retrieval via `/knowledge-mcp`
Cross-Repo Scope	Non-existent; single-file focus	Siloed within individual git repositories	Uniform ecosystem contracts and living decision logs (SimpleAgilityDecisionLog)
Verification Loop	"Looks good to me" visual scan	Human-gated unit test suites	Automated TDD, mutation testing, and machine-readable linter gates
Defect Escape Rate	Extreme; frequent regressions and silent drift	Moderate; bounded by repo-local context	Minimal; deterministic schemas and automated conformance scripts

2. Navigating the Core Disciplines

Core Workflow Disciplines

ClaudeCodeWorkflows — Practical session workflows, slash commands, worktree isolation, and MCP grounding patterns in Claude Code.
ContextEngineeringForCodingAgents — Structuring multi-tiered context: root rules (AGENTS.md, CLAUDE.md), progressive procedural skills (SKILL.md), and token-budgeted external retrieval using Wikantik's Knowledge MCP (assemble_bundle, get_page_for_agent) to eliminate context window rot.
SpecDrivenDevelopment — Transforming informal intent into immutable, machine-verifiable requirements before generating implementation code. Covers EARS requirements, dual-agent handoffs, and living wiki RFCs (SimpleAgilityCommerceDataFlow, SimpleAgilityAttributionRFC).
VerificationLoopsForAgenticCoding — Building test-first gating harnesses, mutation testing, and AST assertions that prevent test-gaming, hallucinated passes, and reward hacking.
StaticAnalysisTools — Equipping agents with machine-readable diagnostics (SARIF, Checkstyle, PMD, SpotBugs, Semgrep) for automated error self-healing.

Scaling & Fleets

SubagentOrchestrationPatterns — Decomposing complex tasks across specialized subagents: planner, implementer, reviewer, and verifier roles with strict tool and model tiering (flash vs pro).
ParallelAgentsAndWorktrees — Mechanics of executing multiple concurrent coding agents on isolated git worktrees without branch collisions or disk contention.
BackgroundAgentsAndAsyncDelegation — Long-running background daemon execution, issue-to-PR pipelines, and asynchronous reactive wakeups.

Extension & Failure Defenses

SkillsVsMcp — The definitive boundary between procedural skills (how to reason and execute steps) and Model Context Protocol (dynamic data plane connectivity to tools and databases).
AgenticCodingFailureModes — The 7 critical failure modes: silent hallucination, reward hacking, context eviction, code bloat, scope creep, prompt injection, and credential leakage — with concrete guardrails.
AgenticCodingMaturityModel — The 5-level operational framework measuring organizational agent maturity.

3. Operational Integration with Wikantik

Wikantik serves as the authoritative, sovereign knowledge backbone for agentic software development across the Simple Agility ecosystem:

Zero-Crosstalk Context: Agents working across separate repositories (stockroom, roller, jakemon) share identical architectural ground truth via /knowledge-mcp without duplicate documentation commits.
Living Architectural Contracts: Normative engineering standards and cross-repo interfaces are maintained as living wiki documents (SimpleAgilityAgentCoordination, SimpleAgilityTelemetryContract).
Actionable Tool Recipes: Consult the AgentCookbook for operational recipes on tool selection (FindingTheRightMcpTool), page verification (VerifyingAnAgentGeneratedPage), and environment setup (ConnectingClaudeCodeToWikantikMcp).