Building engineering teams that ship in the AI era

Build an always-on signal of where the business, competitors, and internal politics are heading. Notice early, not loudly.

Surface the problem whose solution compounds, not the ten that look urgent.

Every yes is an allocation. Saying no well is what separates leaders from order-takers.

Design bets where failure is legible, survivable, and a source of learning.

Kill criteria written in advance are worth more than any retro.

Mix safe-delivery work with uncertain-upside work so the quarter is neither boring nor existential.

The most consequential call an EM makes each quarter.

Explicit starvation beats implicit neglect.

Inference cost and latency are design constraints, not finance problems.

Strategy in the leader's head is strategy the team can't execute.

A version leadership will remember and peers won't undermine.

Trust survives only if the re-narration is honest about what changed and why.

Great engineers are not on job boards. Sourcing is a leadership activity.

Redesigning interviews for the post-generation world.

Rituals that keep the bar where it started, against quarterly pressure.

Offer-to-ramp is where most hires are won or lost.

A practice of hearing what the person cannot say in Slack.

Growth happens outside the job description.

A new developmental arc, and the trap of rewarding volume over judgment.

A structured, time-bound attempt to help recovery before parting becomes inevitable.

When separation is the right call, make it humane, fast, and final.

Tolerating them longer than six months is almost always a mistake.

Conway's law makes topology a first-order leadership decision.

Too many seniors: gridlock. Too many juniors: low ceiling.

If one person leaving breaks the team, the team was built wrong.

The specific practices that make distributed teams work.

Map the actual path a unit of work takes, and where it stalls.

Planning cadence lags delivery cadence when work shrinks to hours.

Design the path explicitly. Don't let engineers improvise it.

What review is for now (taste, architecture, intent), and what to stop reviewing.

Reversibility, observability, blast radius. The criteria for handing off.

The gate is the design. Keep it from becoming rubber-stamp theatre.

Agents fail silently, plausibly, at scale. Monitoring catches them.

Tokens, latency, and reliability budgets are engineering constraints, not afterthoughts.

Documentation as production infrastructure for people and agents.

Each new hire should improve onboarding for the next.

Minimal discipline that outlasts any individual.

Build memory available to both people and agents.

A small number worth defending; ruthlessly audit the rest.

Async is a different social contract, not the absence of meetings.

Where teams compound or stagnate.

What humans still coordinate on when agents run overnight.

Humane rotation design. Burnout produces worse incidents, not better.

Convert a single failure into durable team improvement.

Imagining failure before it happens is cheaper than recovering from it.

Detection, containment, and rollback of agent-driven damage at scale.

Complexity wins by default. Make simplicity a team norm, not a preference.

Debt is a tool, not a sin.

Stake ground and still update when the evidence arrives.

Plausible-reading code is often subtly wrong. Develop the new instinct.

Push back with judgment without blurring the PM role.

The last line of defence against shipping slop.

Voice the concern others are avoiding, without becoming the blocker.

Feature shipped, outcome achieved, learning extracted. Name which 'done' you're optimizing for.

Conventional pyramids underfit non-deterministic systems.

Budget reliability against feature velocity explicitly.

The small set the team will not compromise regardless of pressure.

A learning ritual rather than a gatekeeping one.

A culture document written in reviewer comments.

Say hard things in a way the receiver can act on.

Leaders who can't take critique don't get honest teams.

Four classes a team needs. Focusing on only one produces pathology.

LOC, PR count, keystroke telemetry. Refusing to measure is a leadership act.

Measure outcomes, not activities. Design signal that resists gaming.

Who reads it, what decision it drives, what happens when the number moves.

Detectable weeks before attrition: meeting load, tone shift, response latency.

Concrete signatures: who asks for help, who admits mistakes, who pushes back.

Observable behaviour, not survey score. What you do on Monday.

DORA as baseline, plus what it misses.

A team can hit targets while quietly destroying itself.

When refusing to compare is the leadership move.

A new vocabulary for the work that still requires humans.

Teams that measure shipping without impact drift into irrelevance.

From 'we shipped it' to 'it worked' as the team's self-definition.

Run long-horizon work inside a short-horizon org without losing the team's confidence.

Most major problems were visible six months before they exploded.

The quiet senior, the slipping review latency, the unasked question.

By the time the number moves, the damage is done.

Not enough to ship, but enough to smell the wrong answer.

40 useful hours of attention, not 60. Design the week around it.

Leaders make worse decisions when tired, anxious, or triggered, and almost never notice.

Intervening too fast and too late are both common.

If you can't write it clearly, you haven't thought it clearly.

A decision is an artifact that persists, not a moment in a meeting.

The truest statement of priorities a leader makes.

The first thing that dies under pressure. And the thing most necessary under it.

Most upward cases fail on framing, not substance.

Where most org-wide progress actually happens.

External presence compounds across years. Not a side project. A core asset.

Built in small choices over a decade, in uncelebrated moments.

The disagreement on record, the commitment real.

The personal list, written down before the pressure arrives.

Use power with awareness; push back on peers who misuse theirs.

What's disclosed, what's expected, what's off-limits, before someone crosses a line no one named.