The headline question for software engineers in 2026 is not whether AI can write code. AI can write code. The economically interesting question is which of the engineer's tasks AI can produce at the reliability the team can ship and the error cost the business can absorb. Our v1 model says: none cleanly clear the Replaceable bar, most clear the AI-augmented bar, and the human-critical tail is sturdier than the “will AI replace developers” framing admits.
The tasks we modeled
Software engineering decomposes badly into a single bucket — the actual work mix varies by company, seniority, and team — but our v1 seed covers a representative spread: writing production code against a spec, writing tests, writing documentation, designing system architecture, debugging production issues, code review, on-call triage, and mentoring junior engineers. These are the tasks our v1 capability matrix scores against nine axes.
The cell-level read
Writing production code against a clear spec lands in AI-augmented territory. Capability is high, reliability is decent, but error cost is non-trivial because confidently-wrong code at scale costs production incidents. The substitution-class rules per ADR-016 land this task in the middle band: AI does most of the typing, the engineer owns the decisions, the review, and the rollback plan.
Writing tests and writing documentation are also AI-augmented in our v1 seed — not Replaceable. Capability is high (especially for boilerplate), but the reliability and error-cost gates keep both outside the Replaceable band. A test that passes locally and misses the production edge case carries non-trivial error cost. A doc that confidently misstates an API contract drags down every downstream engineer. The role gets AI assistance on the typing; the engineer still owns the correctness.
Code review — drafting feedback against a diff — also sits in AI-augmented. Capability is high, reliability is mid; error cost varies by the diff (a security-relevant review can be 4 of 5, a styling review is 1). We model the average, which keeps it in the middle band.
Debugging production issues drops sharply into Human-led + AI-assisted. Capability of pattern- matching a stack trace is high; capability of synthesizing “why is this happening only at 2 AM on Tuesdays in this customer’s tenant” is low. The reliability axis is brutal here — AI guesses confidently and is often wrong. Oversight minutes per incident grow. AI accelerates the search but does not own the fix.
System design and architecture lands in Human-led + AI-assisted at the deep end. AI can produce a plausible architecture diagram. AI cannot weigh five years of tech-debt decisions, the team's deployment-confidence curve, and the business's actual scaling trajectory simultaneously. The context axis of irreducible human value scores high; the ambiguity axis scores higher. AI is a sounding board, not the architect.
Mentoring junior engineers is the role's Human-critical task. Trust scores at the top of the irreducible-value scale, context is multi-year, the “why did the senior cut you off in that meeting” conversation cannot be prompt-engineered. AI can answer technical questions; AI cannot be the person a junior engineer trusts with a career question.
Roughly across a typical week
For a mid-to-senior software engineer in our v1 reference role, the baseline distribution across modeled tasks is: zero Replaceable, majority AI-augmented (production code, tests, docs, code review), a meaningful Human-led + AI-assisted band (system design, on-call triage), and a smaller Human-critical tail (mentoring, architecture decisions with multi-year context). The headline pill for the role is AI-augmented territory, but the relevant shape is that the role's mass sits in the middle two classes.
That is the calm-economic read. Most of the week is on the AI- augmented frontier. Part is still human-led. The narrative of “software engineers will be replaced by 2027” is not what the model says — Replaceable is empty for the role in v1 — and the narrative of “AI is overhyped, my job is safe” is also not what the model says.
Where this changes fast
Three axes we will be watching. Reliability is the lever. If the reliability axis on feature-implementation moves from 75 to 85, the cell crosses the Replaceable threshold and the role's share- weighted picture shifts toward 30-35% Replaceable. That is the Klarna-style discontinuity for software engineering.
Oversight minutes are the second lever. Most operational AI cost for software-engineering tasks is reviewer time, not tokens. A meaningful reduction in oversight-per-output (say, from 8 minutes per AI-generated PR to 2) cuts the operational AI cost line nearly 4x. That changes the NPV calculation for org-wide rollouts.
Error-cost configuration is the third. A bank's software engineering has error-cost-5 on most of these tasks; a marketing site's software engineering has error-cost-1. The same capability and reliability scores produce different substitution- class assignments depending on the error-cost configuration. The Wagecard tool lets you override the default for your domain.
What to do with this if you are a software engineer
Three calm-economic moves. First, do the AI-augmented work with AI. That is half your week. Refusing this is leaving productivity on the table for no methodological reason. Second, double down on the Human-critical work. Mentoring, system design with context, on-call triage — these are the axes the irreducible-value cluster keeps protecting. They are also the work that compounds your career. Third, watch the reliability axis. When it shifts, you will want to be the engineer who already understands which of your tasks are affected.
Computing your specific Wagecard takes three minutes. Override the defaults if your role differs (compliance-heavy backend, regulated fintech, safety-critical embedded). The matrix-derived read is at /roles/software-engineer; the live cross-role view by geo × experience is at /insights/software-engineer . Methodology open at /methodology.
The honest read is that 2026 is not the year software engineering gets disrupted top to bottom. It is the year a meaningful chunk of the role's task surface moved into the AI-augmented band, and the rest of the work — the Human-critical part — got more valuable per hour, not less.