The working papers¶
The ideas on this site were developed in the Keenan Vision Working Paper series — three papers, each at the maturity stage its own cover states, none yet published. The digests below carry each paper's argument in plain language, and the papers themselves publish when they are submission-ready.
Why digests, not downloads
Working papers are working documents. Rather than circulate drafts that will change, this page states each paper's thesis honestly, names its status, and points to the plain-language treatment of the same idea elsewhere on this site. When a paper reaches its publication venue, this page will say so.
The Spec Manifest: A Unit of Production for Agentic Software Engineering¶
Vernon Keenan, Ian Gotts & Mehmet Orun — version 4, May 2026. Pre-publication draft.
Every engineering discipline organizes itself around a unit of production — the part, the schematic, the deployable — and the paper's starting observation is that agentic systems have entered production without one. Salesforce's published Agentic Enterprise Architecture, four layers over a trust foundation, is coherent and well-formed; what it does not name is the artifact that carries business intent across those layers. The obvious candidates all fail: the agent is a runtime, the prompt is too unstable to version, the workflow assumes determinism it no longer has, the model is an input owned upstream. The paper then identifies five distinctive properties of agentic software engineering — non-deterministic execution, delegated authorization, compounding provenance debt, cross-surface traversal, and the gap between intent and action — and argues that together they demand a new artifact: the Spec Manifest, a declarative, intent-bearing, provenance-carrying, machine-actionable unit of production. Notably, the paper retracts the program's own earlier position that intent should be a fifth architectural layer. Intent, it argues, is a cross-cutting property that must be infused across all four layers — the same lesson software engineering learned with security and observability. The paper is deliberate about its ancestry: it traces the manifest's prior art through four named traditions — business process modeling (BPMN), formal specification (RFC 2119, Design by Contract), the contemporary spec-driven development movement anchored by GitHub's Spec Kit (with AWS Kiro and Tessl as kin), and governance-as-data (W3C PROV, supply-chain attestation, policy-as-code) — claiming only the intersection none of them occupies. The argument is grounded in the Data 360 Strategy Sprint and the Architecture as Strategy interview study.
The plain-language treatment is the Spec Manifest concept page; how the Spec Kit inheritance was customized for SaaS is Borrowed from software engineering.
Governed Intent Compilation¶
Vernon Keenan, Ian Gotts & Mehmet Orun — working draft, 2026.
The companion paper supplies the discipline the Spec Manifest lives inside. It frames enterprise agentic delivery as a two-stage compilation discipline whose target is not code but a governed operator environment — the assembled context, guardrails, and evidence a worker steps into to do the job. An ahead-of-time stage compiles an engagement's digital twin (its living model of what the customer's business wants) into the Spec Manifest, which plays the role a compiler's intermediate representation plays: the stable, typed artifact between intent and platform. A just-in-time stage then compiles individual work packets at the moment of dispatch, against live conditions. The result is deliberately operator-agnostic: the same compiled environment serves a human engineer or a virtual worker. The paper is candid about what the compiler analogy cannot deliver — because the inputs are human-authored and probabilistic, no formal guarantee exists that output means what input meant, so governance (provenance, decision traces, a maturity ladder) is substituted as the trust instrument. And it confines its novelty claim carefully against four predecessor traditions — model-driven architecture, intent compilation in networking and cloud policy, spec-driven development, and the framing of AI agents as compilers — claiming only the intersection none of them occupies.
The plain-language treatment is the big idea: compiling intent.
The Cognitive DevOps Maturity Model¶
Vernon Keenan — first draft, June 2026; co-authors to be determined.
As large language models absorb the mechanics of writing software, the paper argues, the binding constraint on enterprise delivery shifts from code-generation speed to the governance of intent translation — the reliability with which authorized business intent becomes safe, governed, economically useful software change. It proposes a six-level maturity model for enterprises delivering change with humans, AI agents, workflows, and virtual employees, resting on a System of Intent and graded by the two discriminators SAE J3016 uses for driving automation: the operational design domain a delivery system can own — the conditions it can handle without a human taking over — and who holds fallback responsibility when intent diverges. The evidence is hedged: METR's 2025 randomized trial found experienced developers using AI tooling roughly 19% slower while believing themselves roughly 20% faster — a perception gap the paper calls the era's central diagnostic; DORA's 2024 survey (39,000+ respondents) associated each 25% increase in AI adoption with an estimated ~1.5% decrease in delivery throughput and ~7.2% decrease in delivery stability — associations, not causation; DORA 2025 concluded AI amplifies what a delivery system already is. Two hypotheses are named as hypotheses, with a validation agenda: VE Economics and Quiet Erosion, continued on the Virtual Employee Economics page.
Following the work¶
The papers develop in public view even before they publish: the ongoing analysis stream runs at salesforcedevops.net, and the research program can be reached through keenanvision.net.