14Working as a teamDeep dive ⑥

Onboarding into the agent workflow: 'just use it freely' transfers no craft

What you're good at is tacit knowledge — what to delegate, where to gate — so dropping a newcomer in empty-handed pushes them to one of two extremes: delegate blindly, or not dare to delegate

Read4 min read
Topicsteamwork · onboarding · calibration
TL;DR

What makes you good with an agent — knowing what's worth delegating, where to gate, when to hesitate — is tacit knowledge living in the craft, unspeakable. So "just use the agent freely" transfers nothing: it pushes a newcomer to one of two extremes — delegate everything and leave a mess to clean, or fear it so much they type every line by hand as if it weren't there. The fix is to turn the craft into a teachable path: let them watch first, give a small task with a safety net, then widen as they learn to calibrate the right level of trust.

A new teammate joins. You want them up to speed fast, so you say the breezy thing: "Just use the agent freely, it's great."

A week later, it usually lands one of two ways. Either they trusted it too soon — delegated everything, merged everything, and now you're cleaning a pile of looks-fine-but-wrong in the spots only a practiced hand would suspect. Or the opposite — they fear the agent, won't delegate anything meaningful, type every line by hand as if it never existed. Neither is their fault. It's that your "use it freely" handed them nothing to use it well with.

01What you're good at lives in the craft, not in words

This is easy to underrate. You're good with an agent not because you know a secret tip you can say in one sentence. You're good because of a pile of judgments that hardened into reflex: a glance at a task tells you whether it's worth delegating or faster done yourself, you sense when to set a gate, you smell a suspicious "done."

That pile of judgment is tacit knowledge — it lives in the hands, not on the lips. Ask "how do you know when to trust the agent," and you'll mumble "it... depends, you just learn it by doing." That's honest, but useless to a newcomer: it drops them at exactly where you once stood — having to fall enough times to extract it, except each fall this time is a mess in the shared thing to clean.

The turn is here: a team doesn't get good just because one person is good. That goodness, while it's stuck in one person's hands, is one person's asset. Onboarding isn't passing the newcomer your conclusions — it's building them a path to re-walk the distance you walked, faster and with fewer falls.

02A walkable path, instead of being dropped in empty-handed

What replaces "use it freely" isn't a lecture. It's a three-beat path, each beat widening trust a little — exactly how you learned, just made explicit:

NEWCOMER ARRIVESUSES THE AGENT AT THE RIGHT LEVEL
Beat 1 — Watch: have them sit beside a real session you run, narrating the tacit out loud — "I delegate this because it's reversible," "I gate this because it can't be undone." The craft spoken aloud for the first time.

Beat 2 — Small task with a net: give them a real but reversible task, with the team's shared rulebook as the net. They drive; you review their first merge — not to nitpick, but to point out exactly what to recalibrate.

Beat 3 — Widen gradually: once they can calibrate trust on small work, open to bigger. Trust rises with what they've proven they can do, not with the day they joined.

Notice this path doesn't teach them to "trust the agent" or "don't trust the agent" — it teaches them to <i>calibrate</i> trust per task, across a stretch with a net underneath. What you learned by falling, they learn by a guided road.

03What to pull out of the hands into words

For that path to run, you have to pull a few things out of the craft into something sayable. Not all of it — just the ones a newcomer most slips on if left to fumble alone:

1
What's worth delegating

Say the thing you glance-and-know: what suits the agent, what's faster done yourself. A few real examples from the team beat any abstract principle.

2
Where the gates go

Point out the spots in the team's system that must stop for confirmation — the un-undoable, the shared things. The newcomer has no danger map in their head; lend them yours.

3
The allowed level of trust

Say it plainly: on this team, how far can a newcomer self-merge before a second pair of eyes is needed. An explicit line beats far the alternative of them guessing and guessing wrong.

These three are exactly the tacit knowledge you use without noticing — now said aloud. Saying it once to a newcomer is far cheaper than letting them extract it through a few falls in the shared thing.

04Craft that doesn't transfer dies with the one who holds it

One person brilliant with an agent who can't pass it on is a team's hidden weakness: that brilliance vanishes on exactly the day they take leave, or a fourth person joins and no one had time to guide them. The team thinks it's strong, but that strength hangs on one person.

Hold one line: onboarding into the agent workflow is transferring craft, not granting access. Giving a newcomer permission to use the agent takes a minute; bringing them into the workflow properly takes a path with a net. This is also the team version of probing the agent on a small slice before trusting it on the big task — except this time what's being calibrated is a person, and the guide is you. Your skill only truly becomes the team's asset on the day it no longer needs you to exist.

c
The author

Each story here wraps a lesson paid for in full.

craftagentsomeone building and learning at once

What are you building with agents? Want to trade notes, push back, or build something together — drop a line.

Write hello@craftagent.cloud
58pieces12clustersVI·ENbilingual

Get new pieces by email

Field notes on working with AI agents — occasional, no spam.