How to Decide Remaining SE Scope When One Person Can Ship

Shipping a product by yourself is not the same as performing software engineering. Faster implementation through generators and agents does not eliminate the job.

📑Table of Contents

1. Shipping alone does not shrink software engineering
2. Three jobs that remain when code is abundant
3. Split solo space from multiplayer space
4. Shrink, expand, and remaining-work map
5. Checklist: cut remaining work on the change in front of you
6. FAQ
7. Summary

---

Shipping alone does not shrink software engineering

Typed lines can shrink while the engineered object expands. Routine implementation may move into agents. Design, verification, and operating routines do not move with it.

What remains is fixing intent, controlling generation through architecture, verifying continuously, and separating solo space from multiplayer space.

Place one change on the stack

The action for today is to place one in-flight change onto the Semi-Executable Stack and restore any place where human sign-off has quietly disappeared.

Robert Feldt and colleagues state this in arXiv 2604.15468 (April 2026 write-up of an Agentic Engineering 2026 keynote). They call it a conceptual diagnostic model, not an empirical paper.

The object of engineering extends past executable code into semi-executable artifacts.

• prompts
• workflows
• controls
• organizational routines

---

Six rings of the Semi-Executable Stack

The Semi-Executable Stack has six rings.

1. Executable artifacts (code, tests, schemas, configuration)
2. Instructional artifacts (prompts, natural-language specs)
3. Orchestrated execution (tool use, agent workflows)
4. Control systems (guardrails, evaluation harnesses, escalation)
5. Operating logic (decision prep, role exchange, coordination)
6. Societal and institutional fit (regulation; the paper uses the EU AI Act as external pressure)

---

Four diagnostic questions

Four questions are enough.

• What is the primary artifact?
• Where does meaning depend on interpretation?
• Is the evaluation criterion capability, control, or organizational routine?
• Which human unit absorbs change?

If you cannot pick a ring, ask whether removing that artifact collapses the value.

---

Numbers as supporting color, not a score

The paper’s numbers are supporting color, not a maturity score.

• SPAPI-Tester ran on 100+ APIs and, in production, 193 APIs with 23 failures (22 implementation bugs, 1 document-parse issue)
• A GoNoGo pilot is reported as about two hours saved per decision, with human intervention recommended at Levels 3–4
• Cybernetic Teammate, a 776-person P&G field experiment, reports AI-using individuals matching non-AI teams

Lower creation barriers raise variance and verification demand. Engineering discipline does not become optional.

---

Preserve versus purify

The usable heuristic is preserve versus purify.

• Keep controls that still stop failures
• Cut approval waits that existed only because input used to be slow

Outer rings rest on thinner evidence than inner rings, so do not treat ring numbers as a score. Assign the current ticket one ring. If the work sits outside code, drop generation-speed KPIs.

---

Three jobs that remain when code is abundant

If machines write faster than humans read, construction-centered software engineering is not enough. The remaining core consists of three jobs.

• intent articulation
• architectural control
• continuous verification

Karina Kohl and Luigi Carro argue this in their ICSE-FoSE 2026 position paper (arXiv 2602.04830). Under compression—cheaper implementation, costlier failure—human responsibility polarizes into orchestration and verification.

The oft-cited +55% task completion for GitHub Copilot users (Ziegler et al., 2024) is generation speed. It is not shipping speed.

Reread the classic roles

Reread the classic roles.

• Requirements become continuous intent, not a frozen document
• Architecture becomes the control surface that bounds generation
• Testing becomes the main trust mechanism, not an inspection afterthought
• Maintenance becomes verification across regenerations
• Security is both generation constraints and adversarial checks

---

Accountability collapse and the merge gate

The failure mode is accountability collapse. Weekly regeneration can erase the accident version. You may still reproduce a bug yet remain unable to explain the causal chain.

The paper’s hypothetical dollar loss is not empirical data; do not use it as a statistic. Use the structure: for any module you regenerate, keep logs that can reconstruct the accident version and keep the human refuse points.

---

Merge gate: three points in the repository

This is a position paper, not a maturity model for every organization. The merge gate is still concrete. Before you land a generated diff, require these three points in the repository.

1. an intent statement
2. bounds the change must not cross
3. an executable check

If any of the three is missing, do not merge. The faster generation gets, the more PRs arrive with those three points thin. Adjacent work includes harness design that stops agent failure, and verification that does not leave understanding debt. Both refuse the leap from “wrote quickly” to “safe to ship.”

---

Split solo space from multiplayer space

When AI can 10x anyone, personal speed is the wrong optimization target.

• A terminal or prompt is solo space
• Canvas, review, and ship decisions are multiplayer

Generation may stay solo. Intent and shipping need human time booked first.

Figma’s 2026 collaboration numbers

Figma’s 2026 AI report (full report at figma.com/reports/2026-ai-report) covers 8,403 survey responses and 639 interviews across ten markets.

• People who say AI changed how teams work together rose from 7% two years earlier to 41% this year
• Designer participation in development doubled to 41%
• Developer time on design rose from 44% to 60%
• Seventy-six percent do more than half their work on the canvas
• Ninety percent say design matters as much as or more than before AI; 65% of developers say the same

---

What transfers from the Figma numbers

The report locates friction in the missing shared playbook, not in individual speed, and closes with this: it is not how fast any one person moves, but how everyone moves fast together.

Treat this as Figma’s own research with a canvas-centered product lens. It is not a shipping metric for every software-engineering org. Do not read 41% as “all engineering is now collaborative.” The transferable contrast is solo space versus multiplayer space.

---

One-person payroll, large human graph

Even a one-person payroll sits on a large human graph, argues Meditations on Tech on the solo-maker myth: corpora, models, rented infrastructure, open source. The “99.9% not sole-authored” line is rhetoric, not a survey result.

What remains is deciding.

• what is worth making
• what is correct in this context
• which tradeoff to take
• who the customer is
• whether to ship

Typing is evidence of work, not the work itself. Agents copy good practice in a good codebase and compound rot in a bad one.

---

What the PdEConf official page covers

As discovery context, PdEConf 2026 lists 5 September 2026 at Nakano Central Park Conference, with a public aim of turning specialized skills into product value across role boundaries. That official page covers the event frame, not talk-level numbers. Independent papers and the survey above carry the claims.

Split today’s work and book the multiplayer calendar first.

• solo (generation)
• multiplayer (intent, review, ship)

If you can ship alone but have zero counterpart for translating customer needs, constraints, and tradeoffs, intent ships as a private note. Name one counterpart.

---

Shrink, expand, and remaining-work map

The same fact—one person can ship—forks the optimization target. See which reading your team has fallen into, together with the stop signal.

Shrink, expand, and compression readings

Reading	Optimize for	Remaining human work	Stop signal
Shrink story	Personal generation speed and solo shipping	Prompt wrangling and typing	Raising speed KPIs while review wait grows
Expand story (Feldt)	Quality of semi-executable artifacts (prompts, controls, ops)	Ring diagnosis; preserve vs purify	Treating non-code as “just docs” and leaving it
Compression redefine (Kohl)	Intent, bounds, verification	Refusing generation; explainability; who is accountable	Accident versions disappearing under regen

---

Collaboration and solo-myth readings

Reading	Optimize for	Remaining human work	Stop signal
Collaboration shift (Figma)	How fast the team moves together	Decisions on the canvas; cross-role work	Finishing entirely inside the terminal
Solo-myth undo	Judgment, taste, context	Deciding what is worth building	Reading a single commit author as sole credit

---

Sources for the remaining-work map

Sources as of September 2026:

• Feldt et al., arXiv 2604.15468
• Kohl and Carro, arXiv 2602.04830
• Figma 2026 AI report
• Meditations on Tech

---

How each reading looks in practice

How each reading looks in practice:

• Under the shrink story, review wait looks like slow people
• Under expand, wait is an undesigned control ring
• Under compression, unexplained regeneration is the risk
• Under collaboration shift, terminal-only completion privatizes intent

Role-redesign debates about architects and tech leads overlap this map. The table is not about job titles. It is about what one live change is optimizing today.

---

Checklist: cut remaining work on the change in front of you

Run these seven steps on one in-flight change. The six rings are a diagnosis, not a maturity score. Figma’s 41% is a design-collaboration shift, not a shipping KPI for every SE org.

Diagnose the artifact and the merge gate

1. Pick one in-flight change and write which Feldt ring 1–6 is the primary artifact. If stuck, ask whether removing it collapses the value.
2. If it is ring 1 (code) only, check Kohl’s three points on the PR: intent statement, architectural bounds, executable verification. Missing any of them means do not merge.

---

Separate speed from shipping, and solo from multiplayer

3. Measure generation speed (the +55%-class feeling) separately from shipping speed. If shipping is not moving, drop the speed KPI.
4. Split the workplace into solo (terminal/prompt) and multiplayer (canvas/review/sign-off). Book the multiplayer calendar first.

---

Keep controls, logs, and a counterpart

5. Preserve versus purify: keep controls that still stop failures. Cut approval waits left over from slow input.
6. For regenerated modules, keep logs that can reconstruct the accident version and keep the human refuse points.
7. If you can ship alone but have zero counterpart for customer, constraint, and tradeoff, intent ships as a private note. Name one person.

⚠️ Ring assignment is for discussion. Filling every ring does not mean the team is mature. Spreading AI-driven development across an organization is a later step. Run items 1–4 on one change first.

---

FAQ

If one person can ship, are engineers unnecessary?

Q. If one person can ship a product, are engineers unnecessary?

A. What recedes is part of routine implementation. Feldt reads the object of engineering as expanding. What remains is designing and verifying semi-executable artifacts.

---

If Copilot is 55% faster, is the team 55% faster?

Q. If Copilot is 55% faster, is the team 55% faster?

A. That figure is task completion. Kohl separates generation from shipping and treats unexplained regeneration as risk. If shipping speed is not moving, drop generation-speed KPIs.

---

Is it fine to finish entirely in the terminal?

Q. Is it fine to finish entirely in the terminal?

A. Generation can stay solo. Figma treats the terminal as solo space and the canvas as multiplayer. Intent and ship decisions belong in the latter.

---

Should we implement all six rings?

Q. Should we implement all six rings?

A. They are diagnostic. Do not fill them as a maturity model. Look at the primary-artifact ring and the adjacent ring that would collapse if removed.

---

Can we decide from the conference slides alone?

Q. Can we decide from the conference slides alone?

A. The PdEConf official page covers the 2026-09-05 event frame and the cross-role aim. Claim-level numbers come from independent papers and surveys.

---

Related articles:

• AI-era architect/tech-lead role redesign
• 13588
• 13595

Summary

In an era when one person can ship, remaining work is not typing speed. Widen the scope past code, then cut with verification and collaboration. The next action is checklist items 1–4 on the change in front of you. If you have a ring number, the three PR points, split speed metrics, and multiplayer time booked first, today’s decision is complete.