AI-SDLC
apart from code gen
Agentic delivery changes the artifact, not the typing speed: when the spec, the tests and the review are agent-native and the code is a build product, delivery teams change shape; code generation on its own produces faster commits and the same lead time.
Experiment run, measured result. The only tier that becomes a recommendation.
Confidence
71%human-committedExpiry
36duntil review · 9 Oct 2026Lead time
+6moahead of mainstream awarenessOwnership
SKSam Kowalczykfortnightly cadenceWhere it is
Code generation is mainstream and has not moved delivery lead time in any client we can measure; that was the conviction the field opened on and it has held. Our spec-first run on a live internal repo cut lead time 38% against a code-first control using the same agents, and moved the bottleneck to review, where it now sits. The graveyard holds the obvious next step — merging agent PRs unattended on green CI — because roughly a third of agent-written tests asserted nothing and three regressions escaped in two weeks. The coding harnesses have noticed: two of them shipped the plan as a persisted first-class artifact in Q3, which is the spec-first pattern arriving as tooling. The gate is skills: delivery teams do not yet know how to write a spec an agent can build from, or how to review at the volume agents produce.
Why a Quantium decision hinges on it
Every retail and banking client that bought coding assistants last year is asking why nothing ships faster. The honest answer — that generation was never the constraint — is a delivery-model conversation, not a tooling one, and it is the conversation Accenture and Deloitte are staffing for. If Quantium's delivery pods change shape around spec, test and review before the competitors' do, that is a margin story; if they do not, the coding-assistant spend is a cost line with no return, and the client will notice.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01Spec-first agentic delivery on a live internal repo cut change lead time 38% against a code-first control using the same agents and the same team; defect escape rate was unchanged (x-sdlc-spec-first).
- 02Review time rose from 22% to 51% of lead time under spec-first. The bottleneck moved; it did not disappear.
- 03Agent-written tests caught one regression the human suite missed, and 31% of them asserted nothing meaningful — which is why unattended merge is in the graveyard.
- 01'10× developer productivity.' Commits went up everywhere we looked and lead time did not; the number measures typing.
- 02'The end of the pull request.' Review is the binding constraint and the PR is where review happens; removing it removed the control, not the cost.
- 03Autonomous engineers as headcount replacement. Every run that removed the human reviewer put regressions into production within a fortnight.
- 01A spec format delivery engineers can write in an hour that an agent builds from correctly more than 80% of the time — currently research engineers write the specs.
- 02Agent-written tests with a mutation score above 0.6, so that green CI means something; we measured 0.38.
- 03Review tooling that lets a human review at agent output volume, which nobody has shipped and which is what the harness vendors are now hiring for.
- 01Ship r-sdlc-spec-first as the delivery default: the spec is the artifact, code is a build product, review is staffed as the bottleneck.
- 02Run a Type 3 on review tooling: does a diff-summariser plus mutation-scored tests let one reviewer sustain agent throughput.
- 03Publish pos-agentic-development as the exec framing and keep the tested claims separate from it.
Signals · 10 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
Spec-first run: lead time −38% against code-first control on a live internal repo; review now 51% of lead time
Eight weeks on an internal service repo, two arms with the same agents and team. Spec-first arm: change lead time down 38%, defect escape unchanged, review time up from 22% to 51% of lead time. Agent-written tests scored 0.38 on mutation testing.
extracted claimThe spec, not the code, is where agentic delivery gains come from; review is where they stop.
Accenture and Deloitte AU post 'agentic delivery lead' roles in the same week
Both job descriptions describe reshaping delivery pods around agent output: spec ownership, review capacity, release governance. Neither mentions code generation. Argus inference: competitors are staffing the delivery-model conversation.
Two coding harnesses ship the plan as a persisted, reviewable artifact in the same month
Both now write a plan file before code and keep it as the unit of review. One calls it a spec; the other a plan. Naming event for the field: the method we tested is arriving as a default in the tooling.
extracted claimThe spec is becoming the unit of work in the harnesses themselves.
'The end of the pull request'
Demand-band signal. Argues the PR is obsolete under agent throughput and review should happen at the spec. We agree on where review moves and disagree on removing the PR; kept as the strongest version of the case we do not make.
'Code review is the new bottleneck, and nobody is staffing for it'
Argues that agent throughput moved the constraint to review and that review tooling has not changed since 2015. Matches our mid-run measurement. Widely cited by the harness vendors' own engineers.
'If our developers are 40% faster, why is nothing shipping faster?'
Asked by a retail CTO after a year of coding-assistant licences. Delivery had a tooling answer and the client wanted a delivery-model one. Logged unanswered; became the exec framing for the position paper.
Industry delivery-performance report: AI adoption raised throughput and lowered stability across 3,000 teams
Large survey-plus-telemetry study. AI-assisted teams shipped more changes with a higher change-failure rate and no improvement in lead time. The clearest public evidence that generation is not the constraint. This is also the mainstream-awareness date for the field.
extracted claimAI-assisted teams show higher throughput, lower stability and unchanged lead time.
Logged from Claude Code: agent tests caught a regression the human suite missed — and 30% of them asserted nothing
Product engineer logged from a session on the merchandising assistant repo: an agent-written test caught a date-boundary regression the existing suite did not; a mutation run over the same agent tests found 30% killed no mutants. Tried tier, one repo.
Unattended agent PR merge on green CI: three regressions in two weeks, stopped
Two-week trial letting agent PRs merge on green CI without a human reviewer. Three production regressions, all in code covered by agent-written tests that passed. Became g-autonomous-pr-merge.
speckit — spec-driven development toolkit for coding agents
Open-source toolkit that makes the spec the entry point for an agent build: constitution, spec, plan, tasks. We used its spec format in the experiment's spec-first arm. Star velocity tripled after the harness releases in July.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
Spec-first agentic delivery cuts change lead time by a third or more; code-first use of the same agents does not move lead time.
Review is the binding constraint on agentic delivery; throughput gains stall at the reviewer and review time doubles as a share of lead time.
Unattended merge on green CI fails because roughly a third of agent-written tests assert nothing; green is not evidence.
Harness vendors are moving the plan or spec to a persisted first-class artifact; the spec is becoming the unit of work in the tooling, not just in our method.
Code-generation speedups on their own translate into faster delivery.
Position history · the diff is the product
4 validation runs against a fixed brief. Confidence 42% → 71%.
Experiment concluded: 38% lead-time cut under spec-first, review now half of lead time. Harness vendors shipping persisted plans. Recommendation and position published; skills is the gate.
- Spec-first agentic delivery cuts change lead time by a third or more; code-first use of the same agents does not move lead time.
- Harness vendors are moving the plan or spec to a persisted first-class artifact; the spec is becoming the unit of work in the tooling, not just in our method.
- c-ai-sdlc-2 ↑ 0.6 → 0.76
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
committed · AWChanges how every delivery pod works, including ours.
Timeline
committed · SKPattern is measured on a live repo; the constraint is skills, not capability.
TAM
agent-estimatedAgent-estimated from AU software-delivery services spend exposed to agentic re-pricing. Uncommitted.
Cost
committed · SKA pair for three weeks per experiment; skills work is ongoing and lands on delivery.
Demand
committed · DSEvery retail client with a coding-assistant licence asked the 'why isn't it faster' question this half.
Workforce readiness
agent-estimatedNo delivery pod has written a spec an agent built from without a research engineer. Agent-estimated.
Cost of being wrong
committed · LFWrong delivery model is expensive to unwind mid-engagement; unattended merge is the specific way it goes wrong.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
Retail clients bought coding assistants at scale in 2025 and are the first to ask where the return went. The delivery-model answer is the one they have not heard.
Mechanism · Spec-first pods on the client's internal platforms; review staffed as the bottleneck; lead time measured from telemetry.
Banks cannot run unattended merge under change-control obligations; spec-first with human review at the gate is the version they can adopt.
Mechanism · The spec doubles as the change record; review is the CPS 230 control point.
Large legacy estates where the spec is usually implicit. Spec-first forces it explicit, which is most of the value.
Mechanism · Agent-recovered specs from existing code as the starting artifact; review by the domain owner.
Panel contracts price delivery by role and day. Agentic delivery does not fit the pricing model yet, and agencies cannot buy what the panel cannot describe.
Mechanism · Would apply when a panel arrangement admits outcome-priced delivery.
Red team · the strongest case against
The strongest case against: one internal repo is not a client engagement, and the 38% was measured by the team that wanted it. The review bottleneck we found may be the artefact of a small team with one senior reviewer, and the harness vendors are about to ship the spec-first pattern as a default — at which point the field is a feature, not a delivery model.
- —The experiment ran on our own repo with our own engineers. Client codebases have compliance gates, legacy test suites and reviewers who did not sign up for the study.
- —Review-as-bottleneck could be a team-size effect. With one senior reviewer, anything that raises throughput saturates review; a pod with three would show a different constraint.
- —The 31% empty-test figure comes from one harness on one repo in April. The harnesses have shipped three releases since; the number may already be wrong in either direction.
- —If spec-first becomes a harness default, the differentiator is gone and what remains is change management, which consultancies with 500 delivery staff do better than a lab of twelve.
Source diversity
- Software engineering practice35%
- Vendor / harness20%
- Open-source10%
- Analyst / industry research15%
- Internal / Engel20%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
Agent-written tests that assert something is the precondition for agentic QA; the empty-test finding is that field's starting problem.
Mutation-scored tests and calibrated judges are the same instrument pointed at code instead of models.
If agent-generated volume from citizen developers outruns review capacity, the review bottleneck becomes unmanageable before the delivery model changes.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- SKSam Kowalczyk · Research engineer · SDLC1 drop
- AWAdam Witanowski · Lab Director (acting)1 drop
- DSDev Sharma · Sector owner · Retail1 drop
- OGOllie Grant · Product engineer1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
Agentic delivery works when the spec is the artifact; do not start with the code
strength moderate · 22 citations · review 16 Dec 2026
The direction of agentic development
strength strong · 44 citations · review 18 Aug 2027
Spec-first agentic delivery on a live internal repo
On the firm's internal engagement-tracker repo, features delivered spec-first (human-reviewed spec, agent-written code, human at merge) ship with at least 40% lower review time per PR and no more post-merge defects than the pod's trailing-quarter baseline.
Unattended agent PRs merged on green CI
“Green CI, red incident.” · lived 24 days
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01Does the 38% hold on a client repo with compliance gates and a legacy suite, or only on ours?
- 02What review tooling lets one reviewer sustain agent throughput without becoming the empty-test problem in another form?
- 03When spec-first ships as a harness default, what is the lab's contribution beyond change management?