This technical report, dated May 2026, argues that frontier coding agents able to ingest whole repositories and requirement documents move the bottleneck of software delivery from typing code to the clarity of intent. Its authors read the history of software methodology as a pendulum between rigidity (Waterfall and the Rational Unified Process) and flexibility (Agile and Scrum), and propose Spec-Driven Agentic Development as a synthesis that pursues high documentation fidelity and high execution velocity at once — not, they stress, a return to bureaucratic rigidity.
Much of the report is positional and illustrative rather than empirical. It walks through the classical Waterfall phases with the assistance an agent could give at each, reviews literature on long-context models, LLM refactoring, code migration and AI-augmented testing, and builds its framework, metrics and cost models on that reading; it labels its cost figures as illustrative, non-audited order-of-magnitude anchors.
Key Points
- The report introduces AI-code as a fourth paradigm of software production alongside pro-code, low-code and no-code, defined as code or configuration generated by an AI model or agentic system whatever interface invoked it, and argues that it shifts the governance question from who wrote the code to who specified, reviewed and approved it.
- It proposes the Ambiguity Tax, modelling expected agent hallucination probability as rising exponentially as specification clarity falls, and presents this as consistent with empirical data it cites rather than as a relationship it measured.
- It coins Agentic Linearism for the claimed efficiency of a single uninterrupted synthesis pass from a complete specification, and reframes Waterfall’s requirements freeze and Big Design Up Front (“BDUF 2.0”) as technical optimisations for the agentic execution phase.
- It recasts the engineer as a Spec Architect responsible for intent capture, formal specification, agentic orchestration and adversarial reasoning about agent failure modes, and describes parallel shifts for QA (toward evaluation policy), SRE and platform teams (toward autonomy controls and rollback authority) and product owners (toward intent precision).
- It defines a four-step SDAD workflow — intent capture, formal specification, agentic synthesis, independent verification — in which merge or release authority stays with an accountable human, and depicts it as an SDAD-V model patterned on the classical V-model.
- It proposes Spec Fidelity (completeness, consistency, unambiguity, verifiability) to replace story points, a Synthesis Efficiency Ratio relating a specification’s logic density to tokens spent, an Agentic Autonomy Rate, and a total-cost-of-implementation model with an iteration multiplier; its worked comparison of a Human-Agile and an SDAD module is described as the authors’ own illustrative estimate.
- It closes with hybrid estimation, which pairs planned human effort with a bounded calibration pass, and a staged migration from Waterfall or Scrum through assessment, pilot, hybrid and SDAD-first phases separated by measurable gates.
Notes
The report positions SDAD as a mixed regime rather than a human-absent pipeline, and lists open questions it does not settle: the reliability of AI-generated code at enterprise scale, operational definitions and baselines for Spec Fidelity and the Synthesis Efficiency Ratio, the balance between autonomy and oversight, human–AI collaboration protocols, and when critique among agents improves quality rather than only adding cost. Its future work calls for industry-calibrated measurement of these metrics and longitudinal adoption studies. It relates closely to the broader practice of Spec-driven development and to Agentic Engineering.