Dual-Course Synthesis (Nov 2025 → June 2026)
Domain Status: G1–G10 APPROVED — G1_HARNESS_APPROVED_v1 / G2_TOOL_REGISTRY_LOCKED_v1 / G3_CONTEXT_LAYER_LOCKED_v1 / G4_TOPOLOGY_APPROVED_v1 / G5_EVAL_FRAMEWORK_APPROVED_v1 / G6_VIBE_ENV_LOCKED_v1 / G7_IMPROVEMENT_BOUNDS_v1 / G8_MULTITENANT_APPROVED_v1 / G9_RESEARCH_FLEET_LOCKED_v1 / G10_PRODUCTION_DEPLOY_v1 ✅
A formally-specified Three-Harness Factory Model for autonomous agentic systems, synthesizing the Google AI Agents Whitepaper Series (WP-F1–F5, Nov 2025) and Google Vibe Coding / Agentic Engineering Series (WP-S1–S5, June 2026) into a single, auditable architectural constitution spanning 10 domains (G1–G10).
Agent = Model + Harness
Harness = Context ∪ Constraint ∪ Evaluation
Loop = Read → Write → Test → Observe → Fix
The project's architectural constitution is triangulated across three binding artifacts:
| Artifact | Purpose |
|---|---|
HARNESS_SPEC.md |
Deep architecture — Context/Constraint/Evaluation harness designs, L0–L4 taxonomy mapping, 18-rule constraint catalog, token budgets, traceability matrix |
AGENTS.md |
Global runtime rules — model-routing matrix (Premium/Strong/Flash), HITL gate map (G1–G10), Think–Act–Observe trajectory, always-on static Instructions |
specs/workflow_graph.yaml |
Machine-readable topology — 41 edges, 10 domain nodes, 3 harness nodes, decision overlays, validation invariants |
Course-2 (WP-S) always supersedes Course-1 (WP-F) on overlap.**
| Domain | Focus | Harness Emphasis | Status |
|---|---|---|---|
| G1 | Agent Foundations & Architecture | 🏛️ All three | ✅ APPROVED |
| G2 | Tool Use & MCP | Constraint, Context | ✅ APPROVED |
| G3 | Context Engineering / Sessions / Memory | Context | ✅ COMPLETED (context-v1.0.0) |
| G4 | Multi-Agent Orchestration | All three | ✅ COMPLETED (orchestration-v1.0.0) |
| G5 | Evaluation & Observability | Evaluation | ✅ COMPLETED (eval-v1.0.0) |
| G6 | Vibe Coding → Spec Harness | Constraint, Evaluation | ✅ COMPLETED (vibecoding-v1.0.0) |
| G7 | Self-Improvement (L4 gated) | Evaluation, Constraint | ✅ COMPLETED (self-improvement-v1.0.0) |
| G8 | Multi-Tenant & Policy | Constraint | ✅ COMPLETED (multitenant-v1.0.0) |
| G9 | Autonomous Research Loops | Context, Evaluation | ✅ COMPLETED (research-loop-v1.0.0) |
| G10 | Production AgentOps | All three | ✅ COMPLETED (production-v1.0.0) |
Each domain includes a 6-step delegation runbook (A–F), copy-pasteable Meta-Prompts, and a GIVEN/WHEN/THEN HITL gate.
| Component | Specification |
|---|---|
| OS | WSL2 Ubuntu 24.04 |
| Python | 3.12 (.venv-hermes virtual environment) |
| Harness Runtime | Hermes CLI + Antigravity unified harness |
| Skills Spec | agentskills.io progressive disclosure (L1 → L2 → L3) |
| Model Tiers | Premium Frontier · Strong Coding · Fast Flash (dynamic routing) |
Execution routing: All shell/Python commands route exclusively through the WSL2 substrate:
wsl -d Ubuntu-24.04 bash -c "cd /home/carlospg/workspace/agentic-rd && source .venv-hermes/bin/activate && <command>"agentic-rd/
├── AGENTS.md # Global runtime constitution (always-on Instructions)
├── HARNESS_SPEC.md # Architectural deep-spec (Context/Constraint/Eval)
├── specs/
│ ├── workflow_graph.yaml # Machine-readable factory topology
│ ├── references/ # Immutable whitepaper corpus (WP-F1–F5, WP-S1–S5)
│ └── *.md # Domain blueprints (G2–G10)
├── skills/ # Progressive disclosure skill library (agentskills.io)
├── .gherkin/ # BDD acceptance scenarios (harness + domains)
├── tests/ # Structural & functional test suites
├── logs/ # Agent telemetry & evaluation logs
├── configs/ # Harness configuration (hooks, policy, OTEL stubs)
├── examples/ # Few-shot trajectories & golden patches
├── docs/ # Supplementary documentation & ADRs
├── .gitignore
├── LICENSE # MIT
├── CONTRIBUTING.md
└── README.md # ← this file
- Windows 11 host with WSL2 enabled
- Ubuntu 24.04 WSL2 distribution
- Python 3.12 + project venv
git clone https://github.com/<your-org>/agentic-rd.git
cd agentic-rd
python3 -m venv .venv-hermes
source .venv-hermes/bin/activate
pip install -r requirements.txt # when availableDomain delegation is specified in specs/references/AGENTIC R&D & IMPLEMENTATION BLUE.md. Each domain's Meta-Prompt targets:
- Step A → Premium Frontier (synthesis)
- Step B → Fast Flash (discovery)
- Step C → Strong Coding (decomposition)
- Step D → Strong Coding (scaffold)
- Step E → Strong Coding (tests)
- Step F → Fast Flash (validation + commit)
Every strategic domain gate (G1–G10) surfaces a HARD_STOP decision matrix and requires a human resume token before continuing.
| Field | G1 Foundations | G2 Tools & MCP | G3 Context & Memory | G4 Multi-Agent | G5 Eval & Observability | G6 Vibe→Spec | G7 Self-Improvement | G8 Multi-Tenant | G9 Research Loops | G10 Production AgentOps |
|---|---|---|---|---|---|---|---|---|---|---|
| Status | ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
ACTIVE |
| Resume Token | G1_HARNESS_APPROVED_v1 |
G2_TOOL_REGISTRY_LOCKED_v1 |
G3_CONTEXT_LAYER_LOCKED_v1 |
G4_TOPOLOGY_APPROVED_v1 |
G5_EVAL_FRAMEWORK_APPROVED_v1 |
G6_VIBE_ENV_LOCKED_v1 |
G7_IMPROVEMENT_BOUNDS_v1 |
G8_MULTITENANT_APPROVED_v1 |
G9_RESEARCH_FLEET_LOCKED_v1 |
G10_PRODUCTION_DEPLOY_v1 |
| Tag | — | tool-registry-v1.0.0 |
context-v1.0.0 |
orchestration-v1.0.0 |
eval-v1.0.0 |
vibecoding-v1.0.0 |
self-improvement-v1.0.0 |
multitenant-v1.0.0 |
research-loop-v1.0.0 |
production-v1.0.0 |
| Harness / Substrate | Three-harness Factory + full audit trail | Security Broker Proxy + NPM Pins + Loopback Boundary | Dynamic Token Budget + Honcho Memory + Skills Co-Load Policy | Hierarchical Coordinator + Agent Cards + Policy Seat (DECLARED) | LLM-as-Judge + OTEL Trajectories + Trust Score + Circuit Breaker + 5%/15% Thresholds | Vibe o Agentic Spectrum + SDD + Dune Policy + G5 Inheritance + Slash Command Routing | Bounded Self-Improvement + 10 IT Types + DRAFT/DEBUG/IMPROVE + L4 Gated | ISO-1/2/3 Isolation + SPIFFE SVID + Hybrid Policy Server + Per-Tenant R/B/G | Gherkin BDD Hypotheses + DRAFT/DEBUG/IMPROVE + 7 HITL Gates + Fail-Closed Citation Verification | Spec-driven CI/CD + Canary 1/5/25/100 + Live Policy + OTEL + Doctor + Auto-Rollback |
- WP-F1–F5: Introduction to Agents, Agent Tools & MCP, Context Engineering, Agent Quality, Prototype to Production (Nov 2025)
- WP-S1–S5: The New SDLC with Vibe Coding, Agent Tools & Interoperability, Agent Skills, Agent Security & Evaluation, Spec-Driven Production Grade Development (June 2026)
- Blueprint:
specs/references/AGENTIC R&D & IMPLEMENTATION BLUE.md
MIT — see LICENSE
Generation is solved. Verification, judgment, and direction are the new craft. — WP-S1