Lead coding agents: a product tour
Can this engineer split work across several coding agents and still ship something correct on time?
You’re Mei Tanaka
Candidate
Mei Tanaka leads coding agents to sign outgoing webhooks before a deadline. Then you see what the hiring team sees: the level for each skill, what the rules counted, and every action in order.
10 steps · about 2 minutes
Read the tour as text
Can they split the work so it fits the day?
Mei leads 3 coding agents against a deadline of Tue 18:00. Two of the three work items need the shared signer merged first, so the order matters.
What does each agent get?
Mei makes one decision, then hands each agent its work item, its documents and how it must verify. Billing and notifications start after the signer merges. The record lists exactly what each agent was given, including what it was not: the signer has no key rotation policy, and notifications have neither the signature spec nor the full suite.
What happens when an agent stops and asks?
Without the policy, the signer stops and asks how to rotate keys. If nobody answers within 40 minutes, it goes ahead with its default, a new key only. Mei answers, and when she answered is recorded.
Do they open the diff before they merge?
Mei opens each agent’s diff before merging it. On notifications, where the agent reported that all tests pass, she comments on the line that signs a re-serialized body instead of the raw bytes. Merging without opening the diff is recorded too.
Do they check that it works together?
Mei runs the integration suite on main after the merges, then on the notifications branch, where the customer check fails on 38% of deliveries. She sends the branch back with her comment.
Do they tell the PM what really shipped?
Mei tells the PM what shipped and what did not: the signer and billing shipped, notifications are cut. The status is checked against what actually merged, and claiming work that did not merge is recorded.
What does the candidate see at the end?
In hiring, Mei gets a submission receipt, not levels or feedback, so the answers cannot be passed on.
How is each level set?
In an assessment you publish now, rules set every level from what Mei did, and no model reads the work. Each level says why. Strong needs evidence, such as a merge whose final diff was opened, and doing nothing never reads as strong. Here delegation is at risk because two agents were missing documents the rules say they need.
What did the rules count?
You see what the rules counted: the documents each agent was missing, the question and when it was answered, the comments that landed on a planted defect, and what shipped against what the status said.
Who decides?
A person. The timeline lists every delegation, answer, review, merge and status in plain words, with the time of each, and the actions a level cites are highlighted. You read it and decide.