08 — The hybrid lifecycle
Most institutions won’t pick one band. The common pattern moves a project rightward along the continuum as it proves its worth.
Lifecycle coverage
The continuum measures how much structure surrounds the AI’s work. The lifecycle is the other axis: which phases the AI takes part in. Moving right on the continuum extends AI involvement into more phases.
| Band | Plan | Design | Build | Test | Deploy | Operate |
|---|---|---|---|---|---|---|
| 01Vibe coding | ||||||
| 02AI-assisted | ||||||
| 03Agentic | ||||||
| 04Spec-driven | ||||||
| 05Multi-agent, governed |
In practice, most campus cases stop at band 04. Emerson uses AI for planning, design and building, while people handle review, deployment and operations.
Process models
Agile and waterfall aren’t bands or phases. They are a third, separate axis: how the phases are ordered and how often a team loops back. Any band can run under either model, but AI changes the trade-offs between them.
Waterfall assumed changing code was expensive, so it planned everything up front. Agile assumed change was cheap enough to plan a little and loop often. AI makes code cheaper still, so the bottleneck moves to deciding what to build and checking that it’s right.
Two things happen at once. Upfront thinking comes back as the spec, data model and acceptance criteria, and build loops shrink from two-week sprints to hours.
Emerson’s five steps (vision, data, foundation, experience, operations) run in order like waterfall, with fast iteration inside the experience step. That mix is the pattern now emerging.
Agile taken to an extreme: loops of minutes, with no plan or backlog.
Sit inside existing agile practice. Stories, sprints and review stay the same; the work inside each one moves faster.
Looks like waterfall because the spec comes first, but the spec is living and code regenerates cheaply. Better described as “spec-first, iterate fast.”
Closer to staged gates with continuous delivery inside them, much like regulated agile today.
Large ERP and student-system projects are still governed by waterfall structures: RFPs, phase gates and go-live dates. AI doesn’t remove those. Campus IT shops already using Scrum find sprints shrinking and estimates mattering less. The “definition of done” shifts toward verification.
Most of this comes from practitioner commentary, not measured studies. We found no published higher-ed study comparing agile and waterfall teams that use AI.
DevOps
DevOps isn’t another way to order phases. It’s the automation and shared ownership that make the higher bands safe: automated tests and deployment pipelines, infrastructure defined in code, monitoring, and developers and operations staff jointly owning production. It fills the Deploy and Operate columns of the coverage grid.
Can do without it. Version control is enough for personal and low-stakes work.
Needs automated tests and protected branches at minimum, because agents produce more changes than people can check by eye.
Needs automated checks that verify code against the spec, and a pipeline that deploys the same way every time.
Depends on it entirely: policy enforced in the pipeline, security scans, monitoring and fast rollback.
AI increases the number and size of changes, so delivery and verification become the bottleneck. Google’s DORA research found that a 25% rise in AI adoption went with an estimated 1.5% drop in delivery throughput and a 7.2% drop in delivery stability.
DORA points to fundamentals: small batches and robust testing. AI amplifies whatever practices a team already has. Teams with weak pipelines ship more broken changes, faster.
DORA, Accelerate State of DevOps Report, 2024. Survey-based, modeled correlations, not measured causes.
Its shared environment (sign-on, logging, version control, separate development and production) is a lightweight internal platform. It is what let domain experts ship safely.
Pre-action checks and an audit log are pipeline-style guardrails, applied to agents instead of deployments.
Maturity varies widely. Many ERP and student-system teams still deploy by hand, which in practice caps them at band 03.
Phase by phase
At the structured end of the continuum, spec-driven and multi-agent work, the phases stay the same. What changes is who does the work in each one and what proves it is finished.
Requirements documents, written by hand, that drift out of date
Specs drafted with AI from interviews and tickets, then kept as versioned, testable acceptance criteria
Decide what is worth building and who it serves
Architecture decided in meetings and diagrams
Several design options generated and compared against constraints; decisions recorded as ADRs
Choose the tradeoffs; own security and data boundaries
Developers write most lines
Agents implement scoped tasks in sandboxes; developers pair on the hard parts
Break work into tasks; supply context and conventions
Tests written after the code, coverage uneven
Tests generated from the spec first; evaluations for any feature that uses AI
Judge whether the tests check the right things
Peer review of every diff
An AI does a first review pass; humans review intent, risk and anything unfamiliar
Approve merges; reject code nobody can explain
Release managed by hand or by scripts
CI/CD gates with policy checks; agents write the release notes
Own the go/no-go and rollback
On-call engineers read logs
Agents triage incidents and propose fixes; runbooks become prompts
Make the calls during incidents; run postmortems