AI platform team structure and collaboration

This page explains how teams collaborate on AI features at PostHog. For a high-level overview, see the AI platform overview.

Who does what

The PostHog AI team

PostHog AI Team

PostHog AI Team is responsible for the architecture, performance, and UX/UI of the AI platform. We build and maintain the core infrastructure – the MCP server, skills system, PostHog AI in the web, background sandboxed agents, and shared tooling (`search`, `read_data`, `read_taxonomy`, `enable_mode`). We're also proactive when we see big opportunities for PostHog or when new capabilities can be used across multiple products, like SQL generation or universal filtering.

The PostHog Code team

PostHog Code Team

PostHog Code Team builds PostHog Code, an agent development environment for product engineers. Working with coding agents today is bottlenecked by messy workflows — switching between agents, branches, worktrees, and manually managing PRs across multiple applications. PostHog Code solves this by giving each task its own isolated workspace where an agent works, with everything related to a task in one place instead of scattered across your terminal, editor, and GitHub.

The PostHog Code team owns the desktop app and the task execution pipeline.

The Signals team

Signals Team

Signals Team turns PostHog data into tasks that coding agents can work on — suggested improvements from session replays, fixes for errors from error tracking, new experiments based on product analytics data. Signals surfaces something useful, creates a task with context, and the cloud agent works on it.

Product teams

Product teams own their product's AI capabilities end-to-end. The AI platform is designed so that any team can ship MCP tools and skills independently, without needing the PostHog AI team to be involved. This means you can:

Add MCP tools that expose your product's API to agents
Write skills that teach agents how to accomplish jobs in your domain
Build UI for specific personas using MCP Apps when needed

Once you ship a tool or skill, it's automatically available across every surface – PostHog AI in the web, PostHog Code, Claude Code, Cursor, and any other MCP-compatible agent.

How the teams connect

Together, these teams form the product autonomy loop:

Signals surfaces useful data from PostHog, creates a task with context, and the cloud agent works on it. You review and iterate in PostHog Code.
PostHog AI owns the background sandboxed agents and can start coding agent tasks during chats. These tasks are inspectable in both the web product and PostHog Code.
PostHog Code is where engineers review, guide, and manage agent work across all their tasks in one place.
Product teams ship their own MCP tools and skills independently. Once shipped, these are automatically available across every surface.

Integration vectors for product teams

There are multiple ways product teams can contribute to PostHog's product autonomy vision. These are listed roughly in order of effort, from easiest to most ambitious.

MCP: Expose your APIs to agents

The most obvious and lowest-effort vector. Expose your product's APIs through the MCP server so agents can interact with your features.

Effort: Low

Consumers: PostHog AI, PostHog Code, coding agents (Claude Code, Codex, etc.), Wizard, vibecoding platforms (Lovable, Replit, etc.), ChatGPT & Claude Desktop, and more.

Skills: Teach agents how to do jobs

If you've already exposed your APIs, the next step is explaining how an agent should accomplish typical jobs-to-be-done — analyzing activity in PostHog, debugging why a feature flag was turned off, implementing enterprise features, etc. Skills combine tools, domain knowledge, and step-by-step workflows into templates agents can follow.

Effort: Medium, but the impact is very high.

Consumers: PostHog AI, PostHog Code, coding agents (Claude Code, Codex, etc.), ChatGPT & Claude Desktop, and more.

Signals: Feed the autonomy loop

If your product produces actionable or near-actionable signals — an insight threshold reached, a new error-tracking issue, a frustration pattern detected — use the signals API so agents can discover these hints and act on them later. Signals are what enable the product autonomy loop. PostHog Code acts on plans generated from these signals.

Effort: Low to medium.

Consumers: PostHog Code (local development) and PostHog AI (background agents).

PostHog Code: Features for the agentic development environment

PostHog Code is an agentic development environment where coding agents work on tasks in isolated workspaces. If your product area can make those agents smarter or the engineer's workflow faster, you can build features directly into it. Think PR reviews that check session recordings for regressions, QA steps that verify instrumentation coverage, or task prioritization that weighs your product's signals. This is the highest-effort vector but also the most deeply integrated.

Effort: High.

Consumers: PostHog Code.

Automations & background agents

Run PostHog AI based on triggers from PostHog Workflows, CRON, Temporal, etc., to automate complex workflows. Example use cases: analyze an incoming support ticket based on indexed documentation and respond to the customer, or spawn a new signal like "here is a bug, fix it."

Effort: Medium to high.

Consumers: Your persona using the web browser (UI), PostHog AI, PostHog Code, coding agents (Claude Code, Codex, etc.), Wizard, vibecoding platforms (Lovable, Replit, etc.), ChatGPT & Claude Desktop, and more.

How to get started

The AI platform is self-service by design. Follow the implementation guides to add tools and skills for your product area:

Add MCP tools. Scaffold a YAML definition, enable the operations that make sense, and add a HogQL system table for data access.
Write skills. If your product has jobs that require domain knowledge – specific tool ordering, constraints, query patterns, or reasoning about what data to check – write a skill that teaches agents how to accomplish that job well.
Test with headless agents. Validate that agents can accomplish the workflow by talking to Claude Code or another MCP-compatible agent before building any UI.
Tag the PostHog AI team in PRs. We review PRs that touch the AI platform to ensure they meet our quality bar and integrate well with the rest of the system.

For the full implementation workflow, see Implementing AI features.

When to reach out

You don't need the PostHog AI team to ship tools and skills, but we're always happy to help. Reach out to us in #team-posthog-ai if:

You have an unusual use case that doesn't fit the existing tool or skill patterns
You need something from the AI infrastructure that isn't supported yet
You want design help thinking through how your product should work with agents
You're unsure whether AI is the right approach for a problem – sometimes what seems like an AI problem is better solved another way

Don't hesitate to reach out early, even if it's just a vague idea. We'd rather help you think through the approach upfront than have you discover a dead end after building.

Best practices

Start headless, then UI

Build your product's AI capabilities as headless workflows first – expose the API as MCP tools, write skills for the key jobs. This makes the capability available across all surfaces immediately. Only add dedicated UI when a specific persona needs it. See Implementing AI features for more on this approach.

Start small

Begin with simple tools and iterate based on user feedback. It's better to ship something that works reliably for one workflow than to build something ambitious that works unreliably for ten workflows.

Describe your API fields

API field descriptions flow through the entire pipeline and become what agents read to understand tool parameters. Vague or missing descriptions lead to worse agent behavior. See Adding tools to the MCP server for details.