devdot
← All postsEngineering ·

Devin Bots Can Auto-Merge Low-Risk Changes. Decide What Counts as Low-Risk in Code, Not in a Chat.

Cognition's new Devin Bots manage teams of coding agents for weeks and can auto-merge changes a team marks as low-risk. Before you switch one on, write a risk tier policy by path and enforce it in CODEOWNERS and CI where the bot can't edit it.

Cognition launched Devin Bots on November 17. A Bot is a persistent agent that manages a team of Devin coding agents for days or weeks at a time. It sits in your Slack, remembers context, splits a project into tasks, hands them out, and pings a human when it's stuck. Cognition's example is building the Android version of an iPhone app and keeping it in sync as the iPhone app keeps changing.

The detail that matters most for engineering leads sits further down in Fortune's report. Teams decide which changes need a human review. Low-risk edits, like changing a button's colour, can be approved and merged automatically.

That's a sensible feature. It's also where a lot of teams will get hurt, because almost nobody has a written definition of low-risk.

A button colour is rarely just a button colour

Ask five engineers what a low-risk change is and you'll get five answers. Copy tweaks. CSS. Anything under 20 lines. Anything with tests. Those rules feel fine until you look at what they let through.

A colour change in a shared design token can restyle your checkout's disabled state so the pay button looks clickable when it isn't. A one-line change to a feature flag default can turn on a half-built flow for every user. A small edit to a config file can change a timeout your payment provider depends on. Line count tells you nothing about blast radius.

Humans get away with fuzzy definitions because a human author usually knows which files are scary. An orchestrator running dozens of agents across weeks doesn't carry that tribal knowledge unless you hand it over. Cognition says Bot usage across early-access customers like Decagon, Fireworks AI and Reducto grew tenfold in two weeks. When volume moves like that, whatever your auto-merge rule actually is becomes the policy for hundreds of changes nobody reads.

Define risk by what a change touches

Our rule of thumb: risk belongs to paths and systems. Diff size and the agent's own judgement are poor proxies. Write the rule down where CI can enforce it.

A starting tier list we use with clients:

  • Auto-merge allowed: copy and translation files, docs, isolated components behind a feature flag that's switched off, test-only changes.
  • One human review: shared UI tokens and layout, API handlers, dependency bumps, feature flag defaults.
  • Two reviews, one from the owning team: auth, payments and checkout, database migrations, infrastructure as code, anything touching personal data, and CI config itself.

Then make it mechanical. Put the high-risk paths in CODEOWNERS. Add a CI job that reads the changed file list and labels the PR auto-merge-ok, needs-review or needs-owner. Turn on branch protection so the bot's account can't merge anything without the right label. If a single PR touches paths in two tiers, it takes the stricter one.

The last item in that third tier catches people out. If an agent can edit your workflow files, it can edit the rule that decides whether it needs review. Keep .github/ (or your equivalent) in the strictest tier.

Let the Bot see the rules, but keep them out of its reach

Cognition's own safeguards point the same way. Bots run on separate cloud machines away from engineers' keys, admins choose which repos and tools each Bot can reach, and CEO Scott Wu describes relying on company-set rules "rather than relying on models themselves being smart enough to recognize a risky or malicious prompt." Apply that thinking to merges. Tell the Bot about your tiers so it can plan around them (smaller PRs, fewer mixed-tier changes), but enforce them somewhere it can't edit.

Two more things worth doing before you switch on any agent manager, whether that's Devin Bots, Anthropic's Duet or OpenAI's Dots:

  1. Review the tier list monthly against incidents. If an auto-merged change caused a rollback, the path it touched moves up a tier.
  2. Track auto-merged changes as their own metric. If 80% of a Bot's output skips review, you want to know that was a decision someone made.

Where we'd start this week

Agent managers are going to produce more pull requests than your team can read. That part is fine. The part you can't hand over is deciding which ones need reading, so write that decision into CODEOWNERS and CI now, while the volume is still small enough to tune.

We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!

NEXT POST →Browser Use Made a 15x Pricier Model the Default. Pin Your Model Names in Code.