AI Writes Code Faster Than Your Team Can Review It - Andrew Norrish, Sourcegraph

Highlights

• Sourcegraph builds CodeSearch and CodeGraph products that help owners of complex code bases deeply understand their code.

• Expectations indicate that approximately two-thirds of all code will be written by AI come 2027.

• Developers merge almost twice the number of pull requests with agentic coding, and those pull requests are more than double the size of human-written pull requests.

• Anthropic reported that developer comprehension of code declines as reliance on AI writing the code increases.

• Sourcegraph provides deterministic search across every repository, branch, and commit with a single query.

Timeline

1. 00:00 - 05:00

   The Rise of AI-Driven Development and Code Review Bottlenecks

   Sourcegraph director of field engineering Andrew Norrish outlines how customer mindsets have shifted from asking how many headcount AI can save to managing the overwhelming volume of code produced. As pull request counts and sizes increase, developers prioritize speed over trust just to keep pace with their review queues. Consequently, breaking changes and missing context lead directly to an increase in production incidents.

   • Engineering teams increasingly rely on AI and agentic tools to support their work, causing code bases to balloon.
   • The primary bottleneck in software development has shifted from typing speed to how quickly humans can review AI-written code.
   • AI-written pull requests often lack organizational context and get merged into production without proper human review.
2. 05:00 - 09:20

   Patterns of Failure and the Burden on SREs

   When AI tools lack full codebase visibility, they make guesses that cause cross-service issues and architectural failures. SREs and frontline support bear the brunt of resolving these recurring, costly incidents. Data shows that code review times have doubled, developers find AI-generated code harder to review, and developer comprehension of the underlying code is declining.

   • AI hallucinations occur frequently in large, complex organizations with legacy codebases due to partial environment views.
   • SREs face rising volumes of incidents without runbooks for code that humans have never looked at.
   • Expectations estimate that two-thirds of code will be written by AI by 2027.
3. 09:20 - 13:10

   Where AI Succeeds and the Role of Platform Teams

   Broad automation of software development remains unreliable, but bounded domains such as large-scale migrations and security scanning yield strong results. SRE and platform teams are uniquely positioned to leverage AI safely because they already operate within structured, semi-autonomous processes. These teams build the tooling needed to prevent AI-initiated incidents from running rampant.

   • AI succeeds when its focus narrows to well-bounded domains with necessary context like vulnerability remediations and security scanning.
   • SREs and platform teams act as barometers for AI effectiveness based on the number of incidents produced.
   • Platform teams already work with automated remediation tooling and semi-autonomous runbooks.
4. 13:10 - 18:50

   Sourcegraph Solutions for Investigation, Prevention, and Remediation

   Sourcegraph addresses SRE workflows across investigation, prevention, and remediation. Internal workflows demonstrate how webhook integrations with HackerOne and custom root-cause analysis agents automatically triage security reports and trace production downtime back to specific commits. By providing deep code intelligence and bulk change capabilities, Sourcegraph helps teams maintain stability and compliance despite massive increases in AI-generated code.

   • Sourcegraph provides a full code graph and code intelligence platform that acts as a context engine for agents and developers.
   • Deterministic search and Deep Search allow teams to find root causes across repositories and commits in minutes.
   • Organizations can execute bulk changes across all repositories to remediate risky patterns rapidly.