AI Code Insights: connect production incidents to the code that caused them

What are AI Code Insights?

AI Code Insights is an Obseria capability that links production behaviour to the code changes behind it. When an incident, a new error group or a latency regression appears, it identifies the pull requests and commits that most likely introduced it and shows whether the change was written by a person, by a coding agent or by both together. It turns the release from a black box into a list of concrete, ranked suspects.

The need for this has grown quickly. Coding agents now open pull requests, refactor modules and write tests in many engineering organisations. That raises throughput, but it also raises a new question for engineering leaders and SREs: how does this code behave once it runs in production, and when something breaks, which change was it?

Why linking incidents to code matters

Traditional observability stops at the service boundary. It tells you that the checkout service throws a new exception since the last release, but a release may contain dozens of pull requests from several authors and agents. Someone still has to read diffs, compare stack traces and ask colleagues, often in the middle of an incident.

Closing that last gap shortens resolution time and improves accountability without blame. The right author, human or agent, gets the context needed to fix the issue. Over time, the same data shows patterns: which repositories have a high change failure rate, whether agent-written changes need stricter review in certain areas and where additional tests would have caught a problem before release.

Signals behind AI Code Insights

Attribution combines several signals that Obseria already collects, plus metadata from your source control:

• Release markers and the list of commits and pull requests in each deployment.
• Stack traces and error groups, mapped to the files and lines changed in recent pull requests.
• Trace spans showing which endpoints and code paths regressed in latency or errors.
• Authoring metadata such as commit trailers, bot accounts and pull request labels that mark agent-authored or agent-assisted changes.
• Delivery outcomes like rollbacks, hotfixes and failed release validations.

How Obseria AI Code Insights works

You connect your Git provider and CI/CD system once. From then on, every release in Obseria deployment tracking knows its pull requests. When Obseria Intelligence detects a problem or an incident is declared, AI Code Insights compares the failing stack frames and spans with the changed code, weighs timing and blast radius and ranks the candidate pull requests. The incident shows the top suspects with the evidence for each, and you can open the diff directly.

The same correlation runs continuously in the background, so you get reports on change failure rate, rollbacks and new error groups per repository, team and authoring source. This makes it possible to compare how agent-written and human-written changes perform in your own codebase, based on production facts rather than opinion.

Coding agents can be part of the loop as well. Through the Obseria MCP server, an agent working on a fix can query the errors, traces and SLO impact related to a pull request, and after release Obseria can post a production summary back to the pull request. AI Code Insights costs $10 per user per month; all underlying telemetry follows normal usage-based pricing.

Best practices for AI-written code in production

1. Label agent contributions consistently. Use commit trailers or pull request labels so attribution is reliable.
2. Keep releases small. Fewer pull requests per release make attribution sharper and rollbacks cheaper.
3. Feed production context back. Give developers and agents access to the errors their changes produced.
4. Measure, then adjust review rules. Use change failure rate per area to decide where stricter review is needed.
5. Stay blameless. Attribution exists to speed up fixes and improve systems, not to rank individuals.