Sarthak Garg

When to step back, when to step in

Intervening too fast and too late are both common.

·9 min read·

For a leader, stepping in means acting on work that someone on your team owns: answering a question they were still working through, or taking the task back from them entirely. Stepping back means leaving the work with them, even when you can see a problem in it. The standard advice on choosing between the two is sound, and I follow it:

  • Step in when a mistake would be serious and hard to undo, such as damaged customer data or a commitment the company cannot keep.
  • Leave smaller mistakes alone, because an engineer who finds and fixes a problem learns more than one who is handed the fix, and a team whose leader catches every error stops looking for errors.
  • Match how closely you watch to how much experience the person has with the work.
  • When someone brings you a problem, ask a question before you offer a solution, so that the thinking stays with them. Michael Bungay Stanier built a whole book, The Coaching Habit, around this one.

You can follow all four rules and still not know what to say in tomorrow's design review, because if you follow them together, you end up choosing between two acts, keeping quiet or giving the answer, with the coaching question counted on the quiet side. I think you do a large share of your leading through the question itself, because a question can carry almost none of the answer or nearly all of it, and choosing how much to put in is the decision.

An engineer on the payments team adds a cache in front of the account-balance lookup, and the leader reviewing the design can see that a balance will go stale after a payment is written. The leader has at least four things they could say:

  1. "What would worry you about this once it is in production?" leaves the whole search to the engineer.
  2. "What happens to a cached balance when a payment lands?" points at the right area and leaves the finding to the engineer.
  3. "Would clearing the cache entry on every write fix the stale balance?" names both the problem and the fix, and asks only for agreement.
  4. "Clear the cache entry on every write" is no longer a question at all.

Each version on that list leaves less of the thinking to the engineer than the one before it. The right question moves the person toward the problem without doing the work for them, and more than one wording can do that, so you are aiming at a range of questions rather than one perfect sentence. If you pick a question that carries a little too much or a little too little of the answer, the person still gets closer to the problem, and you can adjust with the next question.

Let the stakes set how far you can go

Start with what happens if the person gets it wrong. When the mistake is easy to undo and touches a small area, the widest question is enough, and so is silence. When the mistake is hard to undo, such as a change to stored data or a gap in security, you can move quickly toward the questions that carry more of the answer, and all the way to telling if time is short. These are the same two tests, whether the change can be undone and how much damage it could do, that you would use before letting an AI agent make a change on its own.

An engineer posts a plan to drop an unused column from the live transactions table during a Tuesday release. The leader skips the open question about what could go wrong and asks directly, "What is the rollback plan if something still reads that column?" The engineer finds a reporting job that does, and the change moves to the following week, which nobody minds.

Start from the person's experience with this task

After the stakes, weigh the person's experience. Andy Grove, who ran Intel, called it task-relevant maturity in his book High Output Management, meaning how much experience a person has with the specific kind of work in front of them. When that experience is low, he said, a leader should give detailed direction and check often; as it grows, the leader can step back until they are checking only the results. Grove tied the maturity to the task rather than to the person, which means a staff engineer with ten years of backend work can be a beginner on a mobile screen, and a junior engineer who has run the release process twenty times may know it better than anyone else on the team, including you. When you choose a question, start wide for someone with deep experience on the task, and closer to the problem for someone new to it. You move that starting point for each person as your trust in them grows.

A senior backend engineer picks up their first Android screen, a list of recent transactions that stutters when the user scrolls. The leader asks, "How many times does this list redraw when one row changes?", which points straight at the rendering cost. A month later, when the same engineer brings a design for a new settlement service, the leader asks only, "What would keep you up at night with this in production?"

Ask a little wider than your instinct

Once you have spotted the problem, you feel a strong urge to say it, and it is easy to see why you give in to it: saying the answer takes ten seconds and feels helpful, while the other person pays for it later, in what they did not learn, where you will not see it. Bungay Stanier calls this urge the advice monster, the part of you that jumps to a fix before you have understood what the other person is dealing with. In Multipliers, a study of leaders who shrink their teams without meaning to, Liz Wiseman describes a type she calls the rescuer: a leader who cannot watch someone struggle and steps in to end the discomfort, which is often the leader's own.

To counter both, ask a question slightly wider than the one that first comes to mind. If the person sees the problem from there, the thinking stayed with them, and if they do not, the next question can carry a little more of the answer. Every one of Bungay Stanier's seven questions is open, which can make a narrower follow-up feel like a failure of coaching, but I think narrowing one step at a time is a normal part of asking well, and kinder than leaving someone to guess.

In a design review for a new checkout flow, the leader asks, "Where could this break?", and the engineer looks at the diagram and says it seems fine. The leader follows with, "What happens to this flow if the call to the card network takes eight seconds?", and the engineer starts sketching a timeout and a retry.

The engineer learns the most in the time between those two questions, while looking at the diagram and not yet seeing the problem, which is why rescuing someone too early wastes a stretch assignment. When you feel the urge to rescue someone, check whether the person needs the help or whether you need to stop feeling uneasy.

Ask early, while a small hint is still enough

Waiting too long is the other common mistake, and every week you wait, the wide questions get less useful. Early in a piece of work, the person has time to follow a wide question: they can get part of it wrong and still recover. As the deadline gets closer, that time runs out, and only a question with the answer already inside it will help before the date. Past a certain point no question helps at all, and you end up writing the plan yourself. The engineer pays for the wait with the fix they never got to make, and you pay by doing the work you had been trying to leave with them.

An engineer estimates three weeks to move the nightly reconciliation jobs to a new scheduler. At the end of the first week nothing has merged, and the standup update says the work is progressing. A question that day, "Which step are you least sure about?", gives the engineer room to name the blocker and solve it. The same gap left until the fifth week ends with the leader writing the plan and the engineer working through a list someone else made.

Spotting a quiet first week like that, when the standup update sounds fine, is a separate skill.

Tell plainly when a question will not help

Some situations need you to skip the questions and state the answer:

  • an incident, where every minute spent on a question is a minute customers wait
  • a deadline that cannot move
  • a security or privacy gap
  • two questions in a row, each carrying more of the answer than the last, that have still not helped

Say what you think should happen and why, so that the person can still learn from the reasoning. A leading question at least shows your answer openly, so the person can argue with it; a question you ask while hiding the answer you want does more damage than telling, because the person has to guess what you think instead of studying the code. After a few questions like that, people start studying you instead of the problem, and every later question feels like a test.

Payments start failing after a deploy on a Friday evening. The leader in the incident channel does not ask what anyone thinks is happening; instead, they say, "Roll back the deploy now, and page the database on-call to check connection counts." The questions about why the deploy passed review wait for the postmortem on Monday.

If the same person keeps needing the narrowest question on work they should know, changing your question will not fix it; diagnosing what is behind the gap comes first, and if the gap is real, a formal plan to close it may follow. When you do have to say something hard, say it in a way the person can act on.

Read what happened and adjust the next question

After the exchange, ask yourself whether the person owned the fix or only carried it out:

  • If they found the problem and chose the fix, the question was about right.
  • If they waited for the next hint, the question was probably too wide, and next time you can narrow sooner.
  • If they agreed and moved on without much thought, it was probably too narrow, and next time you can start wider.

In each case you are judging your question, and you review their work separately.

A leader notices that one engineer has started answering the open question before it is finished. For that engineer, the leader stops preparing pointed follow-ups and starts each review with the widest question available.

None of this needs to be exact, because a small mistake costs little either way: a question that carries slightly too much of the answer costs the person a small chance to learn, and one that carries slightly too little costs a few minutes and a second question, so you can get this wrong for a while without hurting anyone much.

You also get better at it with practice: after enough conversations, you can hear from the way someone describes a problem how much of it they already see, and you pick the right first question more often. A regular 1:1 is a good place to ask people directly whether your hints have been too large or too small.