From 90 Minutes to 10: One Team's ADR Quarter With StudioX

Let me tell you about the Meridian platform team, because their quarter is the clearest before-and-after I have on the ADR Authoring Bot, and because nothing about it required them to change how they work. Eleven engineers, a payments-adjacent product, the usual velocity, and the usual quiet erosion of institutional memory. When I first sat with their lead, Dana, she could name four decisions from the previous year whose rationale nobody could reconstruct. That is not a broken team. That is every team.

The "before" day

Here is a Wednesday from their Q1, before the Mission was in place. A decision to move their event bus from RabbitMQ to Kafka had been reached — not in a meeting, but the way these things actually get decided: a Slack thread on Monday that spilled into a design-review call on Tuesday, resolved in a PR description on Wednesday. Three surfaces, three sets of participants, one real decision. No ADR was written, because writing it meant someone volunteering to stitch those three surfaces into a coherent document while the migration itself was already underway.

Six weeks later a new hire asked the inevitable question — why Kafka, we barely use partitioning? — and the honest answer required Dana to reopen the Slack thread (still there, barely), re-watch twenty minutes of the call recording, and reread the PR. Ninety minutes to reconstruct a decision that had been perfectly clear to five people six weeks earlier. Multiply that by the dozen consequential decisions a team like Meridian makes each quarter, and by every time each one gets re-asked, and the tax is enormous and completely invisible on any dashboard.

Turning it on

We stood up the ADR Mission inside Meridian's own perimeter in an afternoon. The Capture Agent was wired to their GitHub and Slack through instant MCP servers — no integration project, just registering the tools the agent would discover at runtime. The History Agent pointed at an initially empty ADR knowledge base that would fill as records accrued. Nobody on the team was asked to adopt a new ritual. The whole point was that the Mission met them where decisions already happened.

The "after" day

When the Kafka-style decision recurred a month into the pilot — this time a choice to adopt an outbox pattern for a new integration — the shape of the day was different. The decision landed the same way it always had: a thread, a bit of call, a PR. But the Capture Agent had been reading those surfaces, the reasoning core recognized that a decision had crystallized, and by the time Dana opened her portal the next morning there was a drafted ADR waiting in her decision queue.

It was not a wall of transcript. It was structured: the context, the decision, the alternatives that had been floated (a change-data-capture approach, and doing nothing), and the consequences the team had consciously accepted. Every claim was traceable — she could open the observations and see that "we rejected CDC to avoid a new operational dependency" came verbatim from a specific reviewer's comment, not from a model's guess. Because writing to their docs repo is an irreversible action, the Mission had not filed anything; it was waiting on her. She read it, corrected one word, and approved. Ten minutes, most of which was reading.

That is the crux of why the team actually trusted it. The Mission does the reconstruction — the ninety-minute archaeology that no human ever volunteered for — but it never quietly commits to the record of truth. A person approves every ADR through the decision queue before it lands. Ambient capture, human sign-off.

The numbers Dana cared about

By the end of the quarter Meridian had 23 ADRs where the prior year had produced four. The per-record cost had inverted: what used to be an unpaid ninety-minute reconstruction (usually skipped, occasionally done badly) became a ten-minute review of a draft that was already right. Call it two-plus hours saved per decision that previously got documented, and — more importantly — a dozen decisions per quarter that now got documented at all instead of evaporating.

The escapes-avoided column is harder to put a single number on but is where Dana saw the real value. Twice that quarter a proposed change collided with a recorded decision, and the searchable archive surfaced the earlier rationale in seconds — once preventing a re-litigation of the exact RabbitMQ-versus-Kafka debate, and once flagging that a "cleanup" was about to remove a deliberately dual-written path. Each of those is a day of re-argument or a production incident that simply did not happen.

None of this asked the team to be more disciplined. It asked them to click approve. The reasoning behind Meridian's architecture stopped living in the heads of whoever was in the room and became durable enterprise knowledge — the same shift we aim for across the business applications StudioX teams already work in every day.

If you want the leadership framing of why this matters, I make the fuller case in why it matters; and for the architecture underneath that ten-minute review, Trevor lays out the agents and gates in how it works.