Essay · the job after code
Code is the smallest
part now.
The engineers who thrive next won’t be the fastest typists with the best autocomplete. They’ll be the ones who can run a team — theirs just happens to be made of agents. What firms have always looked for in people leaders is now an engineering skill.
Act I · Direct
Know what right looks like.
If you can’t tell good work from bad, you can only rubber-stamp what your agents hand you. Direction comes before delegation.
First principles
Agents produce fluent, confident output whether it’s right or wrong. The only defense is understanding the problem from the ground up — the domain, the constraints, the physics of the system. If you can’t derive what the answer should look like, you can’t catch the moment your team drifts. This is the keystone; every other skill on this page leans on it.
The manager’s margin
The leader who understands the business can direct specialists they couldn’t personally replace. The one who doesn’t gets snowed by every status report.
Owning the outcome
Agents don’t carry pagers. Whatever ships, you shipped it. Product judgment — why the work matters, what failure costs, when a human must stay in the loop — can’t be delegated to something that doesn’t bear consequences.
Accountability is the one thing a manager can never hand down.
Act II · Delegate
Build the team and its environment.
Delegation isn’t handing over a task. It’s designing the conditions where the task can’t go quietly wrong.
Design the org, not just the code
System design now includes deciding what your team of agents looks like: which components and agents exist, what each is responsible for, how work moves between them. Sharp interfaces and precise briefs — schemas, contracts, unambiguous specs — are the job descriptions you write. Vague briefs get confident nonsense back.
The manager’s margin
Org design and role clarity. Most team failures are structure failures.
Harness engineering
Most of what gets listed as separate AI skills — tool design, context quality, retries and timeouts, evals and tracing — is one discipline: engineering the environment your agents work in. Onboarding docs, feedback loops, guardrails, the right tool within reach, honest performance reviews. Metrics, not vibes. This is where the leverage lives.
The harness behind this site is documented in full →
A manager’s output is the system around their people. You can’t buy an agent pizza — you can only make its environment better.
Know your agents’ limits
Delegate what the model is good at; keep what it isn’t; know the difference this month, not last. Limits tell you where to double-check, where to challenge, and where supervision is a waste of your attention.
Knowing your people — who’s ready for the stretch assignment, who needs review on everything, who will confidently accept work they can’t do.
Act III · Verify
Trust is a process, not a feeling.
The failure mode of working with agents is one every manager knows: believing the confident report.
Curiosity to challenge
The best question in the room is still why. Probe the answer, ask for the reasoning, make the design defend itself. Confident output that can’t survive three follow-up questions wasn’t an answer — it was a guess.
The manager’s margin
Good managers don’t take the status update at face value. Probing isn’t distrust; it’s the job.
The tester’s mindset
You’re no longer primarily the builder — you’re the one who decides what verified means. Knowing which tests matter is the skill: a test existing proves nothing about whether the behavior is tested. And bad tests aren’t free — every low-quality test is a tax on every future agent run. Tokens burned, loops slowed, signal diluted.
Acceptance criteria, and inspecting what you expect. A manager who can’t evaluate work ends up managed by it.
Security as risk ownership
Your agents can be socially engineered — that’s what prompt injection is. Input validation, output filtering, least-privilege permissions aren’t compliance chores; they’re the access-control calls any manager makes about who can touch what.
The manager holds the risk the team can’t. You decide what the intern gets prod access to.
Coda
Your team upgrades every quarter.
The one certainty: parts of this page will be wrong within a year. Models change, harnesses change, limits move. Staying agile about the application is table stakes — the rarer skill is staying agile about your own harness and your own beliefs about what AI can’t do. Yesterday’s limit is today’s default setting.
The best managers re-learn their people as they grow. Yours just happen to get dramatically better every few months.
This is a map, not a syllabus.
One engineer’s read on where the craft is going. If you’d draw it differently, argue with me — that’s Act III in practice.