Spec-Driven Development for Agentic SE: why vibe coding hits a team wall

The claim in one paragraph

Software engineering is sliding from AI-assisted “vibe coding” — a tight conversational loop between one developer and a model — into Agentic Software Engineering (ASE), where autonomous agents take goal-level tasks. The paper argues that this is not a smooth upgrade of tooling: it is a socio-technical reconfiguration. Spec-Driven Development (SDD) is proposed as the discipline that makes ASE viable at team scale, with specifications as the canonical, version-controlled contract between humans and agents.

The productivity paradox

Industrial signals are the empirical hook. Faros AI (10k+ developers) reported big gains in individual task completion and merged PRs, alongside much longer review times and more defects. DORA’s AI editions show a similar pattern: throughput can move, instability sticks. METR’s RCT found experienced developers slower on their own repos while believing they were faster. GitClear’s longitudinal data points at less refactoring and more duplication.

The synthesis is familiar systems thinking applied to the SDLC: accelerating code production without expanding review, verification, and coordination capacity just moves the bottleneck downstream — Amdahl’s law for delivery. Vibe coding is the clearest illustration: fast local output, weak traceability and reusable specs.

What SDD changes

Requirements engineering always treated specs as communication and validation artifacts. Under ASE, specs become operational — consumed directly by agents. From them, code, tests, docs, and infrastructure are derived with a clear origin trail. SDD reconstitutes the contracts that vibe coding dissolves: accountability, verifiability, and transferability.

• —
  Harness — technical wrapping of the model plus methodological mechanisms around the team that translate agent capability into team outcomes.
• —
  Human role — from coder-in-the-loop toward orchestrator and verifier, via patterns such as briefing, consultation, review, norm encoding, and orchestration.

How to read it

The field is young; shared vocabulary is still forming. The authors lean on gray literature — vision papers, talks, practitioner reports, tooling docs — and frame the piece as a conceptual foundation and research agenda, not a validated theory. Useful if you are already running coding agents and feel individual speed outrunning team absorption capacity.