Shipping the Agentic Software Delivery Framework
Methodology as repo, not a slide deck
"AI-native" teams can ship unreviewable diffs with no spec trail and call it velocity. Agents widen output; without shared contracts for what to build, how to review, and who merges, you get vibe coding at scale - throughput without a durable system.
Here's what this piece covers: why chat-thread knowledge fails, what commit e55425c (2026-06-18) imported into a canonical repo, what shipped on day one, the named pattern, and three lead habits before you fund "AI adoption."
Part of Field Notes: agentic delivery.
1. The leadership problem
Teams adopt Cursor and Copilot ad hoc. Specs live in tickets. Review norms live in seniors' heads. Merge rules live in Slack. Knowledge evaporates when the thread scrolls away - not versioned, not enforceable, not teachable to the next hire or the next agent session.
Wrong: "We adopted AI."
Right: "Jun 18 e55425c: we open-sourced a delivery framework - spec → agent → gates → draft PR → human merge."
The anti-pattern is vibe coding at scale - agents as a faster keyboard with no methodology attached.
2. What the git history shows
On 2026-06-18, commit e55425c was the initial import of the Agentic Software Delivery framework. Same day, naming and framing tightened: AGENTS.md refocus (7573a39), then the engineering-led reframe (609d567 - next post).
| Artifact | What it holds |
|---|---|
methodology.md |
Progression: project → architecture → specs → implementation → merge |
docs/standards/ + index.yaml |
Engineering standards with a machine index agents consult first |
.agents/skills/ |
Workflow skills: spec-builder, pr-review, qa-testing, gh-stack, worktree, … |
docs/human-in-loop-pr-review-strategy.md |
~780-line merge playbook (article 14) |
templates/AGENTS.md |
Consumer-repo router; AGENTIC_DOCS_ROOT convention |
| Quality gates | Pre-commit hooks, invisible-Unicode lint tooling |
That commit did not prove consumer monorepos were compliant. It proved the delivery system was versioned - a prerequisite for every skill sync that followed.
A tiny shape for "what may an agent touch before human merge?":
PROGRESSION = [
"project", "architecture", "proposals", "features",
"specifications", "gates", "agent_impl", "draft_pr", "human_merge",
]
def agent_may_start(change):
return change.has_implementation_contract and change.entry in PROGRESSION3. What I built
This was not a slide deck. It was a canonical repo meant to be vendored into consumer monorepos.
Methodology progression - one path from intent to merge. Small fixes may enter at specs; uncertain direction starts with proposals. Scope determines entry point - not every change walks the full chain.
Standards library - agents load index.yaml first, then only the standards that apply to the task. Loading the whole corpus wholesale burns context and fails quietly (that pain later drove August compression work).
Skills as workflows - spec-builder drafts contracts; pr-builder packages evidence; pr-review grades against contract and standards; gh-stack / worktree keep lanes isolated. Copy skills without standards and review time is where adoption dies.
Propagation - rsync of methodology, skills, and docs; consumer examples include service and frontend monorepos. One canonical source; forks create drift.
I still treat this as a working system under revision - not a finished doctrine. Naming the thing took four commits on day one. Clarity matters for adoption.
4. Named pattern: methodology as repo
Methodology as repo - invariants, skills, and gates versioned in git - not a workshop deck; the loop is spec → agent → gates → draft PR → human merge.
| Surface | Who uses it | Decision it supports |
|---|---|---|
invariants.md |
Leads, staff+ | What humans retain vs what agents execute |
index.yaml |
Agents | Which standards apply to this task |
| Skills | Engineers + agents | Repeatable workflows without rediscovering process |
| Review strategy | Reviewers | Depth by blast radius, not file count |
What broke: the corpus grew faster than agents could load it. Partial adoption (skills without standards, or standards without merge policy) fails at review time. Measure adoption by merged SPECs and skill invocations - not workshop attendance.
5. What I'd do as a lead
Three habits - steal one this month:
- One canonical methodology repo. Consumer repos sync; they do not fork policy into Confluence.
- Copy skills and standards together. Partial adoption fails when the agent drafts code the reviewer cannot grade against a contract.
- Measure adoption by merged SPECs and skill invocations, not slide-deck attendance.
Further reading
- Field notes: agentic delivery - series intro
- Next: Engineering-led, agent-accelerated
- Framework anchor: commit
e55425c· 2026-06-18
