Insights | 8 West Consulting

The Rediscovery of Deep Specification: How LLMs will Rewrite the Agile Playboo

Written by Brendan Lawlor | 2026 - June 8

The Documentation Dilemma

 

In the modern software development landscape, there is a constant, nagging tension between speed and direction. For over two decades, the "Agile" movement has prioritized rapid delivery and working code over comprehensive documentation. While this shift was a necessary correction to the bureaucratic stagnation of the past, it has left many teams moving at high velocity toward ill-defined destinations. We’ve traded the "analysis paralysis" of the nineties for a modern "directionless drift."

 

However, we are currently witnessing a fundamental shift in this dynamic. We are re-discovering the value of deep software specifications, not because we’ve developed a sudden nostalgia for the slow-moving processes of the past, but because Large Language Models (LLMs) have rewritten the economic laws of technical documentation.

 

This rediscovery is anchored in two interconnected facts:

 

  1. LLMs work better in the presence of deep specifications.
  2. LLMs make deep specifications economically viable.

 

 

LLM's mutually-amplifying qualities when creating and using deep specifications.

 

 

The Ghost of Development Past: From RUP to Agile

 

To understand where we are going, we must respect where we have been. Our current industry standards - the daily stand-ups, the sprints, the Scrum ceremonies - did not emerge in a vacuum. They were a visceral reaction to the "heavyweight" paradigms of the late 20th century, most notably the Rational Unified Process (RUP).

 

RUP was built on the ideal of deep, up-front specification. While the goal was clarity and rigor, the reality was a methodology that buckled under its own weight. We abandoned those deep specifications because they suffered from four fatal flaws:

 

  • Expensive: The human capital and man-hours required to produce them were prohibitive.
  • Perishable: Like fresh produce, they would "go off quickly," becoming outdated the moment the first line of code was written.
  • Untestable: There was no automated way to verify if these natural language artifacts were actually correct.
  • Uncompilable: There was a massive interpretation gap between the written document and the final production code.

 

The Cost of Thinking: Overcoming the "Expensive" Barrier

 

Historically, writing a high-quality specification was an exhausting endeavor. It required "facilitated think-time" - long, expensive sessions where multiple people with different expertise had to be locked in a room together. Because the process required full immersion to maintain context, it effectively prohibited "time-slicing"; you couldn't just dip in and out of deep analysis.

 

LLMs, combined with specialized workflow tools like BMAD, have fundamentally changed this calculation. Textual manipulation like drafting, parsing, and capturing knowledge, is the core strength of these models.

 

By adopting an "agent-as-facilitator" approach, we remove the cognitive tax traditionally paid by the analyst. The LLM manages the conversational flow and administrative overhead, freeing the Subject Matter Expert (SME) from the "fear of the white page." Most importantly, the state of these sessions can be captured and "re-hydrated" at zero cost. This synthesis of agent-facilitation and session re-hydration means we no longer need five people in a room for six hours; we can achieve the same depth through asynchronous, high-intensity bursts of analysis.

 

The Shelf-Life of Ideas: Solving the "Perishable" Problem

 

A recurring grievance among seasoned veterans is that specifications "go off" quickly. In a fast-moving business environment, the deeper and more detailed a specification is, the more perishable it becomes. Any change in business logic or implementation feedback traditionally triggered a cascading manual update process that few teams had the stomach for.

 

LLMs solve this perishability by making document maintenance as inexpensive as the initial creation. We can now iterate on specifications as we move from a Minimum Viable Product (MVP) to later releases. This allows us to "back-propagate" implementation decisions into the specs with ease, ensuring the documentation remains a living, accurate reflection of reality rather than a dusty relic of what we thought we were building.

 

Quality Gates: Turning "Untestable" into Rigorous

 

In the past, it was all too easy for a cohort of analysts to convince themselves of a document's correctness, hand the artifacts to an implementation team, and move on. This disconnect meant developers often inherited unverified, contradictory, and logically flawed blueprints.

 

Modern LLM-based review steps introduce inexpensive, rigorous quality gates:

 

  1. "Grill-me" sessions: Iterative processes where the AI proactively challenges the logic and edge cases of the specification.
  2. Consistency Tests: Automated checks for coherence and logical contradictions within the natural language text.

 

By applying these gates, we establish a new standard for the SDLC: the deeper the specifications, the deeper the tests we can automatically generate and run.

 

The Bridge to Reality: Compiling Specifications

 

The Agile revolutionaries famously insisted, "You can't compile UML" (but oh boy we even tried that). This was a critique of the inevitable interpretation errors that occur when a human developer tries to translate an analyst's artifact into functioning code. This gap often forced teams to effectively start from scratch during construction, which led to one of the core insights of the agile movement - that code is the ultimate design.

 

Unfortunately, that insight has often been used as a license to develop in a directionless way, avoiding up-front analysis even when the application desperately requires it. LLMs now serve as the bridge that spans this gap.

 

We no longer have to choose between up-front specification and iterative development. The LLM brings the scale and precision necessary to "compile" - in a loose but meaningful sense - specifications into actionable plans and code. We can now specify up-front to the full extent the domain allows, while simultaneously iterating across both the specs and the code as new decisions arise.

 

Strategic Implications and Broken Assumptions

 

The assumptions that have governed the Software Development Life Cycle for the last 25 years have fundamentally changed. The problems that Agile and Scrum were created to solve - the high cost and rigidity of planning - now have additional, AI-driven solutions. As we move forward, we must re-examine every facet of our process, from the ceremonies we value to the documentation we have traditionally despised.

 

I believe that very little invention will be required to navigate this new era. Most solutions will be found in the industry's previous attempts to solve these problems. We are not just entering a new era of technology; we are finally gaining the tools to fulfill the promises made during the RUP era, without the "heavyweight" tax that broke it. The future of software engineering isn't about abandoning our past, but finally possessing the intelligence to make its foundational goals a reality.