Applied AI Studio

The Agent Architecture

A domain-agnostic Applied-AI system: an executive problem statement goes in, a funded, governed, production-ready delivery plan comes out — driven by an org chart of specialist agents under one constitution.

Settings / Registry Governance & Orchestration Pipeline (execution) Dev pipeline Human-in-the-loop gate
Layer 1

Settings & Registry

The catalog an IDE reads to install the agents — the top “settings” layer.

🗂️ Agent registry registry/agents.json

All 23 agents, typed orchestrator / pipeline / advisor / dev-pipeline, each mapped to its layer, stage and artifact.

🧭 Stage registry registry/stages.json

The ADLC pipeline: every stage’s input contract, output artifact, consulted advisors, and whether it ends at a human gate.

↓
Layer 2

Governance & Orchestration

Rules everyone obeys, one router, and advisors who advise — but never decide.

📜 Constitution CONSTITUTION.md

Human-final, evidence over assertion, ask before you build, safety for regulated data, one recommended path, artifacts are the interface, full traceability. Overrides every agent.

🧠 Orchestrator orchestrator

Classifies the domain, sequences the stages, consults advisors, enforces the gates. It directs — it never does the specialist work itself.

🏛️ The Administrator = Orchestrator + Human owner  ·  advisors recommend → the administrator decides → reserved decisions stop at a human gate
Advisors — consulted on demand, advise the administrator
model-selector
Task→tier→model + routing/fallback + cost math
cloud-gcp
Vertex AI · Gemini · BigQuery ML · Agent Builder
cloud-aws
Bedrock · SageMaker · Kendra · Step Functions
cloud-azure
AI Foundry · Azure OpenAI · AI Search · ML Studio
cloud-onprem
vLLM · Llama/Mistral/Qwen · Airflow · Milvus
connector-advisor
How to connect Snowflake, Databricks, Bedrock…
domain-advisor
15 industries: regs, data classes, required HITL

The four cloud agents are asked the same question so the administrator compares GCP vs AWS vs Azure vs on-prem like-for-like — then gets one recommended path with the reason the other three lost.

↓
Layer 3

Execution — the ADLC pipeline

Each agent knows one job: its input, its output artifact, its gate. Stage N’s artifact is stage N+1’s input.

GATE
01
intake
Signed PRD
02
process-map
Human/Machine steps
03
assess
Solution verdict + model shortlist
GATE
04
value-prop
Business case / ROI
05
architecture
Ref. arch + model strategy
06
dev-spec
AI Spec + eval gate
07
data-science
Prompts/features + log
08
eval
Golden/adversarial sets
GATE
09
poc-gate
GO / NO-GO
GATE
10
production
Launch-blocker checklist
11
observability
Signals · alerts · cost
GATE
12
brief
Delivery brief + spec handoff
↓  AI Spec handoff — only when approved & funded
Decoupled

The Dev Pipeline

A separate pipeline that consumes the AI Spec. Runs only when the initiative is funded — you can plan without ever writing code.

discovery
Maps the target repo · builds the story backlog from the AI Spec
coder
Implements each story · wires model calls, RAG, tools, eval gate
code-reviewer
The merge gate: spec + guardrails + green eval before merge

Cross-cutting foundations

🔌 Connectors (MCP)

Portable across Claude Code, Cursor & Antigravity. SnowflakeDatabricksBigQueryBedrockSalesforceDatadog — read-only & least-privilege by default; secrets never in the repo.

🏭 Domains (15, pluggable)

Core agents stay domain-agnostic; the domain-advisor plugs in the frame. HealthcareFinancial svcsInsuranceManufacturing+11

📄 Artifacts

Every stage writes a real file (PRD → assessment → business case → AI Spec → evals → POC verdict → brief). The demo shows its work.

🧩 One source, four IDEs

.claude/agents/ is native to Claude Code; AGENTS.md + .cursor/rules/ carry it to Cursor, VS Code & Antigravity.