What should the team be able to change without the builder?

The customer layer: the instructions, the examples and knowledge the agent reads, and the eval rows. Di Santo's team took those three and released changes on their own. The integrations and the runtime stay with the builder, and under the codebase contract so do the harness and the core.

Why keep the go-live eval set with the builder?

Because if the team owns both the change and the check, nothing catches an edit that breaks month three except an undo, and an undo only helps once someone has noticed. The builder keeps the set the agent passed at go-live and reruns it on every change the team ships. That rerun is the reversal rate on the FDE scorecard.

How do you know the handoff worked?

The middle column of the ledger stops saying the builder's name. Time to landing on the FDE scorecard is days from signature to the first change the customer makes without you; Di Santo's three-column check is the same number taken every week. A blank keys line at day 60 means it has not happened.

Who Fixes a Live AI Agent After the Handoff?

Who should fix a live AI agent after it is handed to the team?

THE SHORT ANSWER

The team that uses the agent, through a named person with hours in their week, not the person who built it. Marco Di Santo's October 20, 2026 post describes the failure: one builder held requirements, build, and release, the agent went live fast, and then every correction waited for that builder. Retire that correction queue and replace it with the keys line from Month Two of the FDE Transition, a customer layer that diffs in one command and rolls back in minutes, and a go-live eval set the builder keeps frozen and reruns on every change the team ships. Measure landing with Di Santo's ledger, taken weekly: for the last five corrections, who noticed, who changed it, and how many days passed.

The builder's name in the middle column is the bottleneck, and the fix is a name on the team with hours, not a faster builder. The two sides of this handoff, Di Santo's from inside the customer and mine from the vendor side, agree on the instructions, the examples, the version, and the undo, and differ on who holds the receipt.

The three replacements, the one place I differ with his split, and the weekly ledger are in Kill the Correction Queue: The Agent Lands When the Fix Stops Waiting for Its Builder. The keys line comes from Month Two: The FDE Leaves, the Workflow Has to Stay, and the two numbers it feeds, time to landing and reversal rate, are defined in The FDE Scorecard: Count What Stayed, Not Who Was Sent.

Last reviewed 2026-10-20 · 1 min read

Who should fix a live AI agent after it is handed to the team?

SOURCES

THE LONG VERSION

RELATED ANSWERS