The missing human layer
for AI‑built software
Distill finds the changes that need human judgement, brings in the people who know, and gets them up to speed in a minute.
More code. Not more software.
Agents sped up the inner loop: writing code is cheap. The outer loop, where changes get approved, still runs at human speed. Every change waits for someone, and the queue keeps growing.
+49%
longer from pull request to merge once firms adopted coding agents. Finished work didn’t rise significantly.
Harvard study of 718 firms, 20263h → 9h
to a first human response at Uber, between 2024 and 2026, as agents wrote more of the code.
Uber at AI Engineer World’s Fair10 days
for Spotify to merge 1,000 agent changes, down from three months. They call what’s left “the new bottleneck”.
Spotify at QCon London
So changes get skimmed, and the wrong thing ships.
When nobody can read every line, changes get waved through. The risk isn’t bad code. It’s drift: work that passes every check but isn’t what anyone meant. It comes back as rework.
In the same Harvard study, the share of pull requests sent back for changes nearly doubled.
The people who would catch it never see it.
The person who knows the customer, the spec or the edge case is often outside engineering. Tools suggest engineers from code history. Nobody brings in the product manager.
Widen the outer loop
Today the outer loop is one engineer reading every line. Distill widens it. The safe changes ship straight away, and the few that matter go to the two or three people who know, wherever they sit.
Risk
Know which changes need a human
Distill weighs every change by what it could affect, not just the files it touched. The safe ones ship without waiting, which frees the queue for the ones that matter.
Duolingo lets about 1 in 10 changes ship without a human, and median time to merge fell from 18 to 12 hours. Duolingo at QCon London
PostHog does the same for about 1 in 3 changes to its main codebase. PostHog
Routing
Bring in the people who know
Only a change that needs a human gets a channel. Distill fills it with the people best placed to judge, from what they’ve built, written and discussed, including product, design and the people who know the customer.
Spotify lists finding the best-suited person for a change, even when nobody assigned them, as unsolved. Spotify at QCon London
Understanding
Catch everyone up in a minute
Everyone gets what changed, why, the decisions the agent made and the evidence, without reading a line of code. Deeper only if they want it.
35% of AI-written code reaches production before anyone fully understands it. Undo survey of 300 engineering leaders, 2026
38% of developers say AI-written code takes more effort to check than a colleague’s. Sonar survey of 1,100 developers, 2026
Steering
Point the agent the right way
People question, push back and agree. It all goes straight back to the coding agent, so the change becomes what everyone meant.
Maintainers would have merged only about half of the agent changes that passed every test. The agents never got feedback to act on. METR, 2026
And every change teaches the next. What people agree on becomes guidance your agents follow.
Works where your team already is
No new app to learn. Slack for the team, GitHub and Linear for the work, and whichever coding agents they use.
- GitHub
- Slack
- Linear
- Claude Code
- Cursor
- Codex
Frequently asked questions
Keep your judgement in the loop
Get product updates and early access to Distill.