Framing the MDLC research program
Scope, working definitions and first hypotheses for this repository.
- Author
- 🌱
- Research question
- What will the model development lifecycle look like when agents do most of the building and validation work?
- Why it matters
- Tooling, team structure and governance decisions made now assume a lifecycle that may not survive agentic development.
Executive summary
This note sets up the research program: a seven-stage working lifecycle, a vocabulary for the supporting layers agents depend on, and an initial set of hypotheses to test. It contains no industry findings yet.
Key findings
None yet. This is a framing note.
Evidence
Sculley et al. describe ML-specific technical debt, arguing that the ML code is a small part of a production ML system.[1]
The NIST AI RMF 1.0 defines four core functions: Govern, Map, Measure, Manage.[2]
Industry examples
To be collected in subsequent notes.
Architecture implications
Diagrams in this repository are organized by lifecycle stage and human role, not by technology stack, so that responsibility boundaries stay visible.
MDLC implications
Use seven stages (Require, Plan, Build, Validate, Approve, Deploy, Monitor) as the shared vocabulary across all pages until evidence suggests otherwise.
My interpretation
The scarce resource in an agentic MDLC is human review attention at gates, not engineering effort.
Hypotheses
H1. Evaluation evidence becomes the primary artifact of model development, ahead of code.
H2. Policy, knowledge and skills will be versioned and reviewed like code.
H3. Structured workflows improve outcome reliability more than model upgrades do, at equal cost.
Open questions
Where do classical ML and LLM/agent lifecycles diverge enough to need separate tooling?
Which systems of record will own agent action logs?
What would change my view?
- H1: teams reporting that eval evidence is routinely skipped without consequence.
- H3: benchmarks showing model upgrades dominate workflow changes at equal cost.
References
- D. Sculley et al., “Hidden Technical Debt in Machine Learning Systems,” NeurIPS 2015. papers.nips.ccpaper · S-0001
- NIST, AI RMF 1.0, NIST AI 100-1, January 2023. nist.govstandard · S-0002
Last updated 2026-09-30.