A Size Gate Turns "Keep PRs Small" Into an Enforced Control

Intercom's PR review agent will not approve large pull requests. If a change is too big, too complex, or too broad in scope, the agent flags it and requires it to be broken down before it can pass the gate. That one refusal rule is doing more change-management work than most written policies: with the bulk of PRs agent-driven and a fifth auto-approved with no human in the loop, the size gate is what keeps the auto-approval lane safe (Intercom).

This is a decision gate. It decides whether a change may pass, but it acts on one variable: size. The claim runs in a straight line. Refusing to approve large changes forces small, reviewable, reversible PRs, and small, reviewable, reversible PRs are exactly what change-management frameworks require.

The refusal changes engineer behavior, not just the queue

A policy document that says "keep PRs small" is advice. A reviewer that returns "break this down" before any change can merge is an incentive. Intercom describes the effect in those terms: the strict agent creates "a direct, positive incentive for engineers to ship smaller, more incremental, well-scoped changes."

The mechanism is that the gate prices large changes correctly. A big PR no longer costs its author a reviewer's sigh; it costs a rejection and a decomposition. Engineers respond to that price the way they respond to any other, and the whole team's batch size shrinks. Intercom's stated rationale is the classic small-batch argument made enforceable: "Small changes are easier to review, easier to test, easier to understand, and, critically, easier to roll back when something goes wrong." Their shipping culture held that belief for years. The review agent is what made it self-enforcing. For how a gate is structured as a capability the agent does not have, see A gate is a capability the agent does not have.

Small batches are what the frameworks already require

Change-management controls, SOC 2's CC8.1 among them, expect changes to be authorized, designed, tested, approved, and deployed under a defined process. None of those verbs work on a 4,000-line diff. A change too large to genuinely review is a change the control technically passed and practically skipped.

Scope discipline is the property that makes every downstream control real rather than nominal. A small PR can actually be traced against its stated intent, actually be tested, actually be rolled back. Intercom built the system to this bar deliberately, engaged their auditors (Schellman) before scaling, and confirmed the automated review process and its evidence meet SOC 2, HIPAA, ISO 27001, ISO 42001, and AIUC-1. Their summary is the same argument the Prove It Holds section makes control by control: "When you build for safety, compliance follows."

Enforced small batches reverted 10x less than unenforced human review

Intercom's numbers are the strongest published evidence that a size-gated automated pipeline can improve safety alongside speed:

Roughly 10x fewer reverts is not just a model-quality result; it is a system result. Every change in Intercom's AI lane has passed a gate that refuses anything too large to review properly, then records the decision and the evidence behind it. The published data does not isolate batch size as the only cause, and Intercom does not claim it does. The defensible claim is stronger because it is operational: a size gate gives the reviewer a way to say no before the diff becomes too large to review honestly.

AI volume makes the control load-bearing

Without a size gate, AI makes the problem worse, not better. The 2025 DORA report (nearly 5,000 respondents) found AI adoption continues to have a negative relationship with software delivery stability, and names the mechanism: without robust control systems, "an increase in change volume leads to instability" (DORA). Agents produce working implementations in minutes. If review capacity stays flat, teams either queue up or rubber-stamp, the failure mode Intercom says companies are "drifting into silently." Holding batch size down is what lets an automated approval lane absorb the volume without the rubber-stamping spiral.

Where the published record is thin

Two gaps. First, Intercom is the only company in the three reports that has published a size refusal as an enforced review-gate control; the field is otherwise silent on it, so this page rests on one well-documented system. Second, Intercom does not publish the threshold: "too big, too complex, or too broad in scope" is their description, with no line count or file count attached. Teams adopting the control will have to calibrate their own limit and tighten it as revert data comes in.