PostHog Expands Self-Driving Mode to Auto-Fix Errors With AI Research Agents.

PostHog has detailed how its self-driving automation system handles production errors end-to-end, from signal classification through to opening pull requests that pass continuous integration. The open-source product analytics platform announced last month that it would use AI agents to spot issues and draft fixes. The new writeup describes a five-stage pipeline that ingests trillions of monthly events across analytics, errors, session replays, and experiments, then filters, groups, investigates, and iterates on code changes without human intervention until a pull request is ready for review.

The system starts with an LLM classifier at the top of the funnel. Because production errors often include untrusted user input, an attacker could inject malicious payloads into error messages or stack traces that later get fed to an agent. PostHog's classifier drops signals it flags as suspicious before they reach downstream components. Surviving signals move into a grouping stage that clusters problems by root cause rather than by superficial similarity in text or structure.

Once an issue is promoted for investigation, a research agent retrieves additional context such as session replays, related logs, or experiment data. The agent decides which artifacts are worth fetching rather than ingesting everything up front. If the problem is specific enough, a second agent writes a candidate fix and runs it through CI. The agent iterates on the code until tests pass, then opens a pull request that a human reviews before merging.

PostHog acknowledges that errors make better fuel for automation than vague product signals. A JavaScript exception with a clear stack trace gives the agent enough specificity to propose a fix. A broader observation like "onboarding is broken" usually requires a human to define the problem first. The system is designed to recognize when it lacks the detail to proceed and defer to a person.

The architecture mirrors workflows that software teams already follow manually. Engineers triage alerts, group duplicate reports, pull up logs or replays, write a fix, and wait for CI. PostHog is betting that the same sequence can run in the background with agents handling the repetitive steps and humans approving the output. The company has not disclosed how many pull requests the system has opened in production or what percentage of those PRs merge without further changes.

PostHog's move follows a pattern in observability and incident response tools that use large language models to summarize alerts, suggest root causes, or draft runbooks. The difference here is that the agent not only diagnoses the issue but also writes and tests the fix. That level of autonomy raises questions about code quality, test coverage, and whether CI is thorough enough to catch subtle bugs introduced by an agent.

The company positions the system as a way to reduce time spent in dashboards and logs. Instead of engineers scanning error reports each morning, the agent surfaces issues that have already been investigated and paired with a proposed solution. Teams still review every change, but they start from a draft rather than a blank editor.

PostHog has been public about its AI investments. The platform recently surveyed 9,000 users to understand which prompts drove signups and rewrote its feature flag service in Rust to cut resource use by 95 percent. The self-driving system represents a higher-risk deployment of AI, since it touches production code rather than just summarizing data.

The full technical breakdown is available in a blog post by PostHog engineers Josh and Cleo. The post includes diagrams of the pipeline and discusses tradeoffs around latency, cost, and safety when running agents against live event streams.