agentic-sdlc

The operating model, as code that runs.

A runnable reference implementation of the Agentic SDLC Playbook — seven stages, five planes, a portable artifact chain, an autonomy matrix, and the Substitution Test. Gates that refuse, and evidence that is a by-product rather than a reconstruction.

Follow one change, end to end Watch it run Listen (26 min) The gates that refuse Switch the agent vendor GitHub ↗
12/12 Substitution Test portable — scored from the repository
12/12 Deterministic gates each proven to refuse, not just to pass
24/24 Configuration evals non-interactive regression suite
1 line To change agent vendor Copilot, Claude, Gemini or Codex
0 Models in the gate the decision is arithmetic over policy

The whole argument, on one whiteboard

Hand-drawn whiteboard infographic titled 'The Agentic SDLC: Governing Software in the Age of AI'. Across the top, traditional SDLC and human review as promises and checklists are crossed out and replaced by system verification, drawn as a turnstile a robot must pass. To the right, the portable architecture runs intent, spec, plan and code along a chain into git, labelled an immutable automatic audit trail, beside vendor-neutral standards AGENTS.md and MCP. Lower left, the Substitution Test asks whether you could switch AI vendors over a weekend and have identical audit evidence by Monday morning, next to a productivity J-curve showing AI as an amplifier. Lower right, a table maps planes of architecture to their role and key asset, and an autonomy matrix grids risk against environment.
Click through for full resolution. AI-generated with Google NotebookLM from the playbook text, and one thing in it is wrong: it maps three planes, and §5 has five — Agent Runtime and Evidence are missing. Published with the gap named rather than quietly reproduced. The playbook is the source of truth; this is a way in.

Sixteen seconds, and nothing in it is staged

The nine acts of the control layer, recorded as they ran. The act to watch is the gates refuse — each gate has the thing it protects deliberately broken, and must go red.

A recording of the real session, not a simulated terminal. If a gate had failed during the take, the recording would show it. The CI half is here →

The claim

Most agentic-SDLC material describes a workflow. This describes a control layer, and then runs it. Three things follow from that, and each one is checkable on this site:

Principle 4

No model in the gate

Models diagnose, propose, draft and review. The decision to allow or block is arithmetic over version-controlled YAML — the same tables governance signed off.

Principle 5

Evidence as a by-product

Every gate writes a JSON record as it runs. Nothing is reconstructed at audit time, because a reconstruction is a story about a control, not the control.

Appendix C

Survives a change of vendor

make swap RUNTIME=claude — one line of diff, and the gates, evals and Substitution Test re-score identically under every vendor. Executed, not claimed.

Try it in two minutes

git clone https://github.com/olafkfreund/agentic-sdlc-showcase
cd agentic-sdlc-showcase
python -m venv .venv && .venv/bin/pip install -e '.[dev]'

make build test lint gates     # the closed loop plus the control layer
make substitution              # Appendix C, scored from the repository
make eval                      # 24 configuration regression cases
make negative                  # break each protected thing; watch every gate refuse

make negative is the one that matters.

A gate verified only by passing is indistinguishable from a gate that cannot fail.

The question this exists to answer

Which production changes touched control SEC-API-01, which were agent-authored, at what autonomy tier, and who approved each one?

python scripts/query_evidence.py --control SEC-API-01

Seconds, from the repository. The playbook calls answering that in minutes rather than a week the single highest-value output of the whole programme. It works because every artifact in the chain opens with a machine-readable header, and every gate emits a record keyed to a control id.

Where to go next

Start here

One change, end to end

A user story from the sentence someone said in a meeting to a signed artifact — and what stays identical when you change vendor.

90 days

Two organisations adopt this

Step by step at a tier-1 bank and a growth-stage payments firm — and why everything that differs between them lives in four YAML files.

16 seconds

Watch it run

Two recordings of the real session — the control layer, and the same control layer running in CI. Nothing re-typed, simulated or spliced.

Stage by stage

The seven stages

What each stage's control point is, and which workflow enforces it.

The control layer

The gates that refuse

Twelve deliberate violations and the gate that stops each one.

Appendix C

The Substitution Test

Twelve checks executed against the tree — no marks for a document that claims.

Portability

Switching the agent vendor

Copilot, Claude, Gemini, Codex — one command, and the proof that nothing else moved.

Source

The playbook itself

The full v1.1 text this repository implements, published as it is on disk.

Demo data. The evidence records, attestations and change ids here are produced by this repository's own pipeline against a synthetic payments service. They are not any institution's audit records, and nothing here should be presented as one.