PostHog Ships Autoresearch Mode to Automate Code Optimization With AI Agents.

PostHog has shipped Autoresearch mode, a new capability in PostHog Code that turns metric optimization into an autonomous process. Instead of manually testing changes to improve a performance target, teams can now spawn a cloud task that handles the entire feedback loop: instrumenting the codebase, establishing a baseline, running experiments, and keeping only the improvements.

The feature builds on PostHog's existing Code product, which already offers AI-driven coding assistance for analytics workflows. Autoresearch extends that foundation by letting users define a target metric, such as page load time, conversion rate, or error frequency, then handing execution over to an agent that iterates in a closed loop. Each iteration tests a change, measures the impact, and decides whether to commit or roll back. The result is a hands-off approach to optimization that runs continuously in the background.

Users select a metric they want to maximize or minimize. The system automatically instruments the codebase to capture that metric, establishes a performance baseline, then begins experimenting. At each iteration, the agent proposes a code change, deploys it, measures the result, and compares it to the baseline. If the change improves the metric, it is retained. If it degrades performance or shows no improvement, it is discarded and the loop continues.

The approach mirrors A/B testing but removes the manual overhead. Instead of setting up experiments, waiting for statistical significance, and manually applying winning variants, the entire process runs autonomously. The agent is effectively a research loop that learns which changes move the needle and which do not.

PostHog engineer Fernando Gomes announced the feature on LinkedIn, crediting the PostHog Code infrastructure that made the feature possible. The company has been expanding its AI-driven development tools over the past year, previously launching self-driving mode to auto-fix errors with research agents and building integrations that surface analytics inside developer workflows.

Autoresearch is aimed at teams that already track key metrics but lack the time or resources to run structured optimization experiments. Startups optimizing for activation rates, SaaS teams improving page speed, or product engineers reducing error rates can spin up a task and let the system iterate while they focus on other work.

The feature also fits into a broader trend of agentic AI tools that automate repetitive, high-frequency tasks. Instead of relying on developers to hypothesize, test, and implement changes manually, Autoresearch shifts that burden to a system that can run dozens or hundreds of experiments faster than a human team.

Autonomous experimentation introduces risks. If the agent optimizes for the wrong metric or makes changes that improve one metric at the expense of another, teams could end up with unintended side effects. PostHog has not yet detailed safeguards such as approval workflows, rollback limits, or multi-metric constraints, which would help mitigate those risks in production environments.

The feature also assumes teams have a well-defined metric worth optimizing. If the target is too narrow or poorly chosen, the system may improve a number that does not translate to user value. And because the loop runs autonomously, teams need visibility into what changes are being tested and why, which may require additional logging or review processes.

Still, for teams already using PostHog's analytics and experimentation tools, Autoresearch offers a way to turn passive measurement into active optimization without adding headcount or process overhead. It is a practical next step for product teams that want to iterate faster on performance and conversion metrics, following PostHog's broader push into AI-assisted development workflows.

Autoresearch mode is now available in PostHog Code. Teams can access it through the PostHog dashboard and configure it to target any metric their codebase already tracks.