Example · Software Developer (early career)
Tickets, bugs and tests are getting cheap. What holds is senior work nobody has handed you yet.
18 of your 40 hours still need you
- 9 stay yours
- 9 faster with AI
- 22 tools could take
A typical week at the middle pace. Your plan uses your own tasks.
Is software engineering a dying career?
By our task-by-task count, 18 of a typical 40-hour week as a software developer still need a person. Tools could already take about 22 hours. If the parts of the job you find hardest are vague asks from internal users and 2 a.m. data mysteries on call, those are the two most durable parts of your job. If watching an agent do your afternoon in twenty minutes worries you, that concern is fair too, and it applies to about 56% of your week.
Read more
Both are true at once.
Anthropic's March 2026 labor-market paper puts computer programmers at the top of all occupations for observed AI exposure, at 75% task coverage. Stanford's "Canaries" data through June 2026 shows employment of 22-to-25-year-olds in exposed occupations 19% below trend, driven by reduced hiring, not layoffs. If you are early in your career, you are in that cohort. The exposure depends on which kind of coding. Well-specified tickets in a codebase the model can read are the most exposed. Your job has a large slice of that and a smaller slice no model can do without you: the parts in people's heads, in production data, and in who signs off. The plan moves your week from the first slice to the second in 90 days.
The hard truth
Your CS degree now comes with a harder first search.
A computer science degree used to feel like a sure thing. Right now, recent computer science and computer engineering grads are out of work more often than grads in almost any other major.
Sources (2)
- Recent grads in computer engineering had 7.8% unemployment and computer science 7.0%, among the four highest of 74 majors. New York Fed, Labor Market for Recent College Graduates (ACS 2024 data), 2026-02
- Mid-career computer science grads still earn a $120,000 median wage. New York Fed, Labor Market for Recent College Graduates, 2026-02
What stays yours
You know how one tricky internal system actually works (permissions, billing, whatever people come to you about)
That knowledge lives in a few bug fixes, some chat threads and your head instead of the repo. A model reading the code sees what it does, but it misses why and which customers break if it changes.
You go find the person who filed the ticket and ask what they meant
Turning a vague ask into a precise one happens before any agent can start, and becomes more of the job as the build step gets cheap.
You have been on call through real incidents
You have seen the system fail in ways nobody wrote down. Saying "roll back now" carries accountability no tool will be handed soon.
You have been near live data migrations
Irreversible operations on customer data are where checking is hardest and someone must answer for the result. It is the least exposed corner of software work.
You know which internal users to trust about requirements
Time spent learning that "ops says X, means Y" is context a new hire (or a model) rebuilds slowly.
What's next for a software developer?
Plan A is your best move inside the work you do now. Most sessions in the 13-week plan build toward it.
- A
Become the engineer of record for one internal system
Why it fits
Why you. If you are already, informally, the one who knows a system, going from "the one who knows" to "the owner" is documentation, alert routing, and one conversation with your manager.
The gap. You have no system doc and have not run a migration end to end. Your manager may not see you as an owner yet, because you have never asked.
The trade-off. Internal tools are low-visibility. Ownership reaches a promo packet only if you make it visible, which is a separate skill. Owning a system also means owning its pages.
Other moves that fit
These are also good bets in this field. Read them if Plan A doesn't suit you.
Become the person who can say "this change is safe"
Why it fits
Why you. When writing gets cheap, the constraint becomes "who can tell whether it's right." Time spent receiving review, plus your on-call shifts, is the raw material.
The gap. Your tests cover your own changes but not a legacy system's behavior before an agent touches it. You have no migration checklist, and you review for style more than invariants.
The trade-off. If you want to keep building, this means reading more than writing, and some days it feels like proofreading for a machine. On the other hand, spec-writing is still building, one layer up.
Move toward the customer boundary (implementation engineering)
Why it fits
Why you. Implementation engineers translate between customer operations and the product: "go ask what they meant," aimed outward. If vague asks bother you, that discomfort points to a skill you have not named yet.
The gap. You have not been on customer calls yet, and you do not know the implementation team's tools.
The trade-off. Pay can drift away from engineering pay over five years, and you give up some options toward senior individual contributor roles. Treat it as the hedge rather than the plan.
What's a plan B for developers?
Plan B is a job AI is creating that these skills already fit. One session a week in the plan trains you for it, so it is ready if your job shrinks or you want a change.
BPlan B, picked for this role
AI deployment strategist
Going to find the person who filed the ticket and asking what they meant is how AI deployments succeed with real customers.
Other options
AI workflow builder
You can already code, so building and shipping working agent automations for an internal team is reachable within months.
AI evaluation specialist
Being on call through real incidents teaches you how systems fail, which is the instinct needed to test AI features before release.
In your own plan you can switch to any of the 47 roles.
The first two weeks
The first 2 of 13 weeks. Three sessions a week of 45 minutes to 2 hours, about 3 to 4 hours in all, plus one 40-minute Plan B session. By week 13 you'll have a runbook and invariants doc for your system, an agent guidelines file with before-and-after results, and a one-page summary of what you own.
- Week 1: Measure your week
What you do
- ALog your time for one week. Match each block to a task in this report's task list.
- AList every system you have touched in this job. Star the ones people ask you about.
- AUse Claude Code, Anthropic's coding agent, or your company's pilot tool to build one small feature on a side project. Write down its three worst mistakes.
- BWork through Claude Code 101, a free Anthropic course, on a small personal or open-source repo.
- Week 2: Write down what you know
What you do
- AWrite down the invariants, the rules that must always hold, for the system you know best. For example: "role X can never see Y." Aim for at least 8.
- ASpend 30 minutes with a senior engineer. Show them the doc and ask what is wrong.
- ARead Google's guide How to Do a Code Review. Re-review your last five merged PRs (pull requests) against it.
- BTake Anthropic's free course AI Capabilities and Limitations. Note which agent failures you would catch in code review and which you would miss.
Three courses from the plan
Checked 2026-09-22.
- How to Do a Code Review (eng-practices)Google · Free
Yes, week 2. These are the best two hours here.
- PagerDuty Incident Response DocumentationPagerDuty · Free, Apache 2.0
Yes, before your next on-call.
- Site Reliability Engineering / SRE WorkbookGoogle · Free online
Read selected chapters only.
Risks beyond AI: moderateHiring-cycle risk
Interest rates and the 2022–23 hiring hangover explain part of the entry-level slowdown. That changes how urgent this is, but the direction stays the same. At a small company, ownership and promotion also depend on very few people. If they say no, the path changes.
Where this could be wrong
- The shares are guesses until week 1. If your log shows much more feature work, more of your week is work AI can do, and becoming the person who can say a change is safe becomes urgent.
- The scenario may be too slow or too fast. If agents get supervised production access within 18 months, the work AI only helps with today becomes work it can do alone, and owning a system is all that holds. If adoption stalls, you have five years, not three.
- As its authors say, the Stanford figure is descriptive rather than causal. Some of the 19% is interest rates and the 2022–23 hiring hangover. That changes the urgency but keeps the direction.
- We may be over-weighting the system you know best. If it is scheduled for a rewrite or a vendor replacement, own a different one instead. Ask in week 3.
Also matched: software developer, software engineer, junior developer. Research checked 2026-09-22. Page updated .