Seven Disciplines Become Eight

Part 7 closed out the machinery both frameworks share — the ladder, the evidence columns, the calibration against an unreliable witness. This part is the hinge of the series, because it answers the question the mirroring has been begging since part 1. When agents started writing most of the code and the human's job shifted to directing the work, the tempting organisational move was to bolt a row onto the existing competency matrix — “AI tooling, rate yourself 1 to 5” — and call the model updated. I built a second framework instead, and the claim this part defends is the one the next part inherits: the AI era does not retire the software disciplines — it transforms them, and an honest accounting of the transformation turns seven disciplines into eight.

A map, not a memorial

Neither framework is a translation table for the other — each stands alone — but every discipline in the traditional model has a traceable fate in the AI one, and tracing it is the best argument for why the new model looks the way it does. Here is the whole map at a glance; the rest of the part walks it.

Three kinds of fate, then: disciplines that split, disciplines that transferred whole, and disciplines that dissolved into sub-disciplines of something new. Plus one absentee the map does not show, which gets its own section.

The split — one craft becomes four disciplines

Engineering Craft carried 25 points as the technical quarter of the traditional model, and it does not survive as a single row. It fissions along its own sub-discipline lines, and each fragment lands with a new name.

Writing code becomes describing outcomes. The centre of the old craft — the ability to produce the implementation — transforms into Specification Engineering (22): the ability to describe an outcome with enough precision and completeness that an autonomous system can produce the implementation for you. Programming skill becomes specification skill; the typing was never the hard part, and the new framework finally prices it that way.

Software architecture becomes orchestration. Understanding how the pieces of a system fit together and where the joints are is precisely the skill Agent Architecture & Orchestration (12) measures — except the pieces are now planning agents, executing agents, handoffs, checkpoints and recovery paths rather than services and queues.

Debugging becomes evaluation design. The old craft asked “why is my code wrong”; Evaluation & Quality Engineering (10) asks the prior question, “how would I even know whether this output is right” — eval harnesses, holdout scenarios, regression detection across model versions. Diagnosis shifts from reading stack traces to designing the instruments that make wrongness visible at all.

Security expands rather than converts. The old Security sub-discipline maps into AI Safety, Security & Governance (8), stretched over attack surfaces that did not previously exist — prompt injection, data exfiltration through generated output, bias in the artefacts an agent produces.

Add the fragments up and 52 of the new framework's 100 points descend, wholly or primarily, from the old 25. The technical core did not shrink in the transformation — it more than doubled its share of the model while ceasing to be about writing code. That single arithmetic fact is, to me, the strongest justification for building a second framework instead of adding a row.

The transfers — judgement keeps its price

Two disciplines cross the map with their weights intact, and both are disciplines of judgement rather than production.

Leadership (15) becomes Intent Engineering (15). What Leadership priced in the traditional model was encoded judgement at increasing scale — decisions made well, alignment driven, delegation handled. Intent Engineering prices the same commodity delivered to a different audience. Driving alignment among humans becomes encoding alignment for agents: value hierarchies that say which objective wins when two conflict, decision boundaries an agent may not cross alone, delegation and escalation frameworks that operationalise “ask before you act” into structure. The institutional judgement a senior human absorbs through years of osmosis has to be handed to an agent on day one, in writing — and the practitioners good at articulating it are, in my experience, the same people who were good at the human version. Not everything transfers — mentoring stays with humans. But when I set Intent Engineering's weight, I was pricing the same scarce thing Leadership's 15 always priced, and the equal number is the mirroring made visible.

Business Acumen & Strategy (7) becomes Domain Translation & Systems Thinking (7). This is the discipline where the skills transfer with the least transformation of all: product thinking, trade-off reasoning, and the unglamorous institutional knowledge that makes systems make sense — knowing why the billing module carries that one exception for Canadian customers, and getting it written down where an agent can consume it. The weight stays at 7 for the same reason it was 7 before: it is a meta-skill that amplifies the other disciplines rather than producing value alone. Same rationale, same number, new consumer.

The dissolved — three disciplines lose their rows

Three traditional disciplines keep none of their nameplates, yet almost all of their content — redistributed into sub-disciplines of the new model.

Delivery (18) dissolves into the apex. Work breakdown becomes Problem Decomposition & Task Design inside Specification Engineering; incremental value delivery becomes the requirement that subtasks be independently executable and independently testable; dealing with ambiguity becomes the discipline of self-contained problem statements, where every hidden assumption must be surfaced because there is no daily stand-up at which an agent can ask. Specification Engineering, in other words, has two parents — Engineering Craft supplied the describing, Delivery supplied the decomposing — which is part of why it outweighs everything else.

Feedback (13) dissolves downward and sideways. Effective communication — the old discipline's heaviest sub-skill — becomes Instruction Design & Clarity, which is not incidentally the heaviest sub-discipline of Prompt Craft: writing an unambiguous instruction is technical communication with a non-human reader. Knowledge sharing becomes Knowledge Architecture & Curation inside Context Engineering: the team wiki grew up into the information environment an agent boots from.

Self Organisation (10) inverts. Reliability used to be a human virtue — being the person nobody has to chase. In agent systems, reliability is an engineered property: checkpoints, error recovery, graceful degradation, the machinery that keeps a long-running workflow productive when a step fails at two in the morning. Economic thinking survives more literally than anything else in this discipline — it becomes reasoning about token spend, model selection and the cost structure of orchestration, where using an expensive model for routine work is the new over-engineering.

And Collaboration (12) is the honest gap in the map: it stays human. There is no agent-era counterpart to building relationships and handling disagreement with colleagues, and I refused to invent one for symmetry's sake. This is the clearest evidence that the eight-discipline framework is a counterpart to the seven, not a replacement for it — it measures the human-directing-agents job, not the human-among-humans one. The engineer still needs both scorecards, which is exactly the “one person, two scorecards” arrangement part 1 promised.

The missing fifth discipline

Readers who came to this series from my learning material will now do the count and hunt for something. That material teaches a five-discipline stack — Prompt Craft, Context Engineering, Harness Engineering, Intent Engineering, Specification Engineering — and that stack is the conceptual spine of the competency framework. Four of its five names appear among the eight disciplines. Harness Engineering does not.

Let me say it plainly, because leaving it implicit would cost readers real confusion: Harness Engineering is not one of the eight disciplines — it is absorbed into Context Engineering and Agent Architecture & Orchestration. The information side of a harness — the persistent files, memory and conversation state a project accumulates so an agent starts every session already knowing things — is measured under Context Engineering, chiefly its Memory & Conversation State Management sub-discipline. The machinery side — the tools an agent may invoke, the protocols and connections it reaches through, the automation wrapped around its sessions — is measured under Agent Architecture & Orchestration, chiefly Agent Tooling & Protocol Design. A harness is where information meets machinery, and as a teaching concept that junction absolutely earns its own layer of the stack. As an assessable competency, its two halves are better measured where they land. Nothing was dropped in the move from five learning disciplines to eight assessed ones; one concept was reclassified into two.

The genuinely new

Which leaves the parts of the map with no ancestor at all. Context Engineering (18) is the clearest invention: nothing in the traditional framework contemplated curating an information environment for a colleague who reads everything you give it and remembers nothing between sessions. The nearest human analogue — onboarding documentation — was a corner of Feedback's knowledge sharing, not an 18-point discipline with retrieval design and window strategy inside it. Prompt Craft (8) is new in the opposite direction: a skill born already commoditised, the table stakes of the era rather than its differentiator.

Why exactly eight, and not twelve? Because in this model a discipline must earn three things: a defensible weight, a five-level progression with observable evidence at every rung, and a failure mode you can point to when it is missing. Candidates that could not carry all three became sub-disciplines instead — which is precisely what happened to the harness.

So the transformation is a refactoring, not an extension: one discipline split four ways, two transferred at their old price, three dissolved into the fabric, one stayed behind with the humans, and two rows are genuinely new. What the refactor cannot settle by itself is whether the new weights are right — and the two most conspicuous numbers in the model are the apex and the foundation. Next, the defence: why Specification weighs 22 and Prompt Craft 8.