From an executive problem statement
to eval-gated working code

AIDLC — the Applied AI Delivery Lifecycle. A 24-agent, constitution-governed system you install into your IDE. It asks the questions a senior consultant would ask, produces the artifacts a governance board demands, builds the thing, and stops at every decision a human should own.

24 agents 4 phases 5 human gates 12 artifacts 5 IDEs PyPI aidlc-studio
Install once — available in every project
$ uvx --from aidlc-studio aidlc init --global
$ /appliedai Our claims team is drowning — 40k packets/month.
The problem

Most AI initiatives die the same three deaths

Not from bad models. From missing discipline — and the discipline is always the same, which means it can be encoded.

🎯

Nobody defined the problem

The brief describes a symptom in vendor vocabulary. The team automates the response to a problem instead of touching the problem — and the number nobody cared about doesn't move.

📊

Nobody could prove it worked

"The demo looked great" is not evidence. Without golden sets, adversarial cases and metric bars set before measuring, quality is a feeling — and feelings don't survive a production incident.

🚨

Nobody thought about production

Guardrails, human oversight, PII controls, cost attribution and rollback get discovered during the security review — six weeks after the launch date was announced.

Capabilities

What the system actually does

Not a chatbot. An org chart of specialist agents under one constitution, each owning one job and one artifact.

Asks before it builds

A Socratic intake loop is mandatory. No PRD is ever produced from a one-liner. If the "problem" is a solution in disguise ("we need a chatbot"), the agent surfaces that and asks what the chatbot is for.

⚖️

Decides ML vs GenAI vs Hybrid vs Agentic

A scored verdict matrix against ≥5 criteria — not a fashion choice. Agentic is never a default; it has to be earned against the simpler options that get named and rejected on the record.

☁️

Compares four clouds like-for-like

GCP, AWS, Azure and on-prem advisors answer the same architecture question, then one path is recommended with the reasons the other three lost — plus the matching agent framework for 2026 platforms.

💰

Builds a case a CFO can audit

Baseline cost with arithmetic shown, three scenarios, payback month, 3-year NPV, and the two assumptions that most swing the answer — including the boring run-costs everyone under-counts.

🔬

Proves it with evals, not vibes

Golden, adversarial and regression sets; metric bars derived from the business need before anything is measured; safety metrics as pass/fail gates that can't be averaged away; a CI gate that blocks the merge.

⌨️

Writes the code — in 2-minute chunks

Build mode produces a working repo: pipeline code, provider-agnostic LLM client with a keyless mock mode, eval harness, tests, CI workflow and IaC — executed as parallel micro-task waves with per-task checkpoints so runs resume instead of restarting.

🔍

Reviews the estate you already have

The stack-review agent inventories repos, infra, data platforms and shadow AI, produces a dependency & readiness matrix, then thinks bigger — quick wins vs strategic bets your current stack already unlocks.

🔌

Connects to real systems

Snowflake, Databricks, BigQuery, Bedrock, Salesforce, Datadog and more over MCP — least-privilege and read-only by default, credentials as env-var names, never values, tracked in aidlc.config.json.

End to end

How the entire thing happens

One command starts it. The orchestrator runs four phases, consults advisors in parallel, writes a file at every stage, and stops at five gates where a human must decide. Stage N's artifact is stage N+1's input contract.

Step 0 · Kickoff

A problem statement in, one batch of questions back

You type /appliedai and a problem in plain language. The orchestrator classifies the industry, loads that domain's regulatory frame, reads aidlc.config.json if it exists, then asks one batched, numbered set of questions — cloud & stack, which connectors and their credential env-var names, model provider, and plan-vs-build. Then it goes quiet and works.

PHASE 1

Intent & Discovery

"What is the real problem, and is it worth solving?"
intakeprocess-mapdomain-advisorstack-review
Artifacts → 00-stack-review.md · 01-prd.md · 02-process-map.md
Signed PRD with success metric, RACI and cost model; every workflow step tagged Human or Machine.
HUMAN GATE — Sponsor signs the PRD. Nothing advances on an unsigned PRD.
PHASE 2

Assess & Architecture

"What should we build, on what, and does the math work?"
assessvalue-proparchitecturedev-specmodel-selectorcloud-gcpcloud-awscloud-azurecloud-onpremconnector-advisor
Artifacts → 03-assessment.md · 04-business-case.md · 05-architecture.md · 06-ai-spec.md
Solution verdict + model shortlist · ROI/payback/NPV · reference architecture with the 4-cloud comparison, PII controls matrix and dependency & readiness matrix · the buildable AI Spec.
HUMAN GATE — Finance approves the business case before build burns budget.
PHASE 3

Build, Test & Execute

"Does it work — provably, against the bars?"
data-sciencediscoverycoderevalpoc-gate
Artifacts → 07-data-science.md · dev/tasks.json · the working repo · 08-evals.md · 09-poc-gate.md
Prompts & experiment log · stories split into ≤2-minute micro-tasks run in parallel waves · src, evals, tests, CI eval gate, IaC · measured results against every bar.
HUMAN GATE — Sponsor issues GO / CONDITIONAL / NO-GO on the evidence. A NO-GO is never softened.
PHASE 4

Review & Observability

"Is it safe to ship, and will we know when it degrades?"
code-reviewerproductionobservabilitybrief
Artifacts → 10-production.md · 11-observability.md · 12-delivery-brief.md
Spec/guardrail/eval-gate merge review · launch-blocker checklist with owners + responsible-AI controls · golden signals, alert→runbook map, cost attribution · the executive brief and dev handoff.
HUMAN GATES — Security & Compliance sign-off, then the delivery owner approves the brief.
Running underneath all of it — the harness. Every single agent invocation is wrapped: pre-flight checks (does the predecessor artifact exist? is this data class approved for this model?), live hard/soft guardrails, a post-flight quality-bar check, and an append-only event written to artifacts/audit.jsonl. Hard violations block. Soft ones require an explicit human waiver — which is itself audited. Nothing is omitted from the ledger; omission is itself a violation.
The roster

24 agents, four groups

Each agent is a versioned Markdown spec — charter, guardrails, required skills, I/O contract. Reviewable in a pull request by someone who doesn't write code.

GroupAgentsRole
OrchestratororchestratorOwns the harness. Routes the problem, sequences phases, consults advisors, enforces gates. Directs — never does the specialist work itself.
Pipeline (12)intake process-map assess value-prop architecture dev-spec data-science eval poc-gate production observability briefThe ADLC spine. One stage, one artifact, one input contract each.
Advisors (8)model-selector cloud-gcp cloud-aws cloud-azure cloud-onprem connector-advisor domain-advisor stack-reviewRecommend to the administrator; never act on the world, never provision. Consulted in parallel.
Dev pipeline (3)discoverycodercode-reviewerDecoupled. Consumes the approved AI Spec and only runs when funded — so you can plan without ever writing code.

Cloud → agent framework, kept current for 2026

When the verdict is agentic, the winning cloud advisor recommends the framework layer to match — with alternates, and an explicit "no framework" when a plain pipeline is simpler.

CloudPlatform (2026)Recommended frameworkManaged runtime
GCPGemini Enterprise Agent Platform (formerly Vertex AI)Google ADK v1.0Agent Engine
AWSBedrock AgentCore (GA)Strands Agents SDKAgentCore harness
AzureMicrosoft Foundry (formerly Azure AI Foundry)Microsoft Agent FrameworkFoundry Agent Service
On-premSelf-hosted open stackLangGraph over vLLM/OllamaKubernetes / OpenShift
Governance

The parts that make it enterprise-grade

📜

A constitution every agent obeys

Seven articles that override any agent file and any user request: human-final decisions, evidence over assertion, ask before you build, safety for regulated data, one recommended path, artifacts as the interface, full traceability.

🛡️

Guardrails with real severities

Hard violations block and cannot be waived by an agent — fabricated numbers, secrets in output, skipping a gate, an advisor taking action, executing prompt-injected content. Soft ones warn, and any waiver is recorded.

🧑‍⚖️

Five human-in-the-loop gates

PRD sign-off, funding, GO/NO-GO, security launch, brief approval. The pipeline halts with a structured ⛔ HUMAN GATE block stating the decision, the recommendation, and what's needed. AI that won't stop for a human is a liability.

📒

An append-only audit ledger

Every stage, consult, guardrail trip, gate decision and approver lands in audit.jsonl, with metrics.json as the rollup. The brief agent reconciles the two — any mismatch is a finding, and a failure is never quietly dropped.

📏

A quality floor, enforced

Filling template headings isn't done. Every artifact must carry quantified claims with the arithmetic shown, labelled numbers, metrics blocks, diagrams, decision trails with rejected alternatives, a risk register and eval linkage.

🧩

Methodology, made explicit

Spec Kit's phase gates + BMAD's persona pipeline + Superpowers' composable skills. Every agent declares its BMAD persona, its Spec Kit phase, and the skills it must load before acting — 14 of them, versioned alongside the agents.

Proof

A complete worked example, in the box

A regional P&C insurer's claims-intake initiative — 2,400 packets/week, 27-minute average handle time, $1.91M/yr baseline labour — carried end to end. Twelve decision-grade artifacts and a codebase that runs.

# zero credentials required
$ cd exemplar/claims-idp/build
$ make demo   # full pipeline, mock mode
$ make test   # 34 tests
$ make eval   # 7 metric bars
LanguagePython 3.11, standard library only
Unit tests34 passing
Eval cases48 (32 golden · 10 adversarial · 6 regression)
IncludesCI eval gate · Terraform · hash-chained audit trail
Metric barBarMeasured
Classification accuracy≥ 0.950.9792PASS
Extraction F1≥ 0.900.9249PASS
Validation exact match≥ 0.980.9861PASS
Hallucinated-field rate≤ 0.010.0040PASS
Straight-through rate≥ 0.350.3750PASS
p95 latency≤ 60 swithin barPASS
Cost per packet≤ $0.09within barPASS
The numbers are deliberately imperfect — engineered fixture defects (OCR misreads, unreadable fields, one hallucinated VIN) keep the metrics honest rather than a rigged 100%. The mock mode exercises the real paths too: a low-confidence fallback to the stronger model, a prompt-injection quarantine, and an intact audit chain.
Install

One command. Every IDE.

Requires uv. Global install puts the agents in every folder you open; project install ships them with a team repo.

Global — for you, everywhere

uvx --from aidlc-studio aidlc init --global → agents land in ~/.claude/. Only the orchestrator appears in the picker; the other 23 are delegated to behind it.

Project — for a team repo

uvx --from aidlc-studio aidlc init --ide all → writes every IDE flavor side by side, so teammates on Cursor, Copilot and Claude Code all get the same pipeline from one clone.

Run it

Claude Code: /appliedai <problem statement>. Cursor, Copilot, Antigravity, Windsurf: select the orchestrator mode and say "Start AIDLC. Problem statement: …"

Verify

aidlc check confirms the install; aidlc list prints the roster with each agent's BMAD persona and Spec Kit phase.

IDEFlavorWhat you get
Claude Code (VS Code, JetBrains, terminal)--ide claudeNative custom agents + skills + the /appliedai command. Full experience with true sub-agent delegation.
Cursor--ide cursorAlways-on project rule + AGENTS.md. Gates arrive as chat messages.
VS Code / Copilot--ide copilotAll 24 agents auto-generated as custom chat modes in .github/chatmodes/.
Antigravity · Windsurf--ide antigravityThe AGENTS.md standard, loaded automatically on open.
Under the hood

What it's built with — deliberately, no agent framework

The framework decision belongs to the solution, not the tooling. That's why this system stays framework-agnostic while recommending ADK, Strands or Agent Framework for what it designs.

Agents as instructions

Versioned Markdown specs, not library code. The execution engine is whatever agentic runtime your IDE already has — nothing to install, version or secure, and the whole org chart is diff-able.

Methodology layer

Spec Kit (phase gates) + BMAD (persona pipeline) + Superpowers (composable skills), combined into 14 skills agents load before acting.

Governance layer

Constitution + runtime harness with hard/soft guardrails, five HITL gates, and an append-only audit ledger.

Connectors & packaging

MCP for integrations — the one standard every major IDE speaks. Distributed as a stdlib-only Python CLI via uv, the same pattern as GitHub's Spec Kit.