Pineapple on pizza

No data available

Feature ownership
Members

Goals

Q4 2026 objectives

Objective 1: Reduce the complexity of building AI products (Driver: Charles Vien)

Motivation: We should be the core that powers every agentic product at PostHog, but some teams have built their own systems instead of using ours. Two gateways are still in production and product-specific logic in tasks makes every new integration slow.

What we'll ship:

  • Move all products fully onto the new gateway, sunset the legacy one, and remove custom provider keys, so all model traffic goes through one gateway (Driver: Brandon Leung)
  • Ship one task interface any team can build on, with a well defined contract for what a task is, what state it's in, and who can access it (Driver: Charles Vien)
  • Ship a configurable router with provider policy per product or team (Driver: Brandon Leung)

Objective 2: Every AI unit of work is costed and attributed (Driver: Brandon Leung)

Motivation: Product teams top ask is accurate task costs and spending controls. Today the gateway ledger and Django credits are separate systems and we only find abuse after the bill arrives.

What we'll ship:

  • Know and cap what every run and generation costs: exact LLM and compute cost per run, caps at mint, product-level rate limits, and live budgets (Driver: Brandon Leung)
  • Agree on a default billing model for the whole runtime that connects the gateway ledger to Django credits, written up as a pricing review RFC (Driver: Ben White)

Objective 3: Safe model and sandbox access (Driver: Alessandro Pogliaghi)

Motivation: Security has been neglected and we aren't talking enough about abuse, data exfiltration, and prompt injection.

What we'll ship:

  • Ship security at the compute and inference layers: an egress proxy for deterministic rules and trace investigation for the fuzzy cases (Driver: Alessandro Pogliaghi)
  • Move the self-driving pipeline off gVisor onto VMs and cut agent server boot time with a warm-pool control plane (Driver: Alessandro Pogliaghi)
  • Catch abuse before the bill arrives, with fraud checks at signup, per-user caps, and spend-velocity alerts (Driver: Brandon Leung)

Objective 4: A reliable agent runtime teams can build on (Driver: Jonathan Mieloo)

Motivation: People see sandboxes and tasks as slow and unreliable and building on or debugging them is painful.

What we'll ship:

  • Harden and simplify the agent runtime package, deciding which harnesses product teams need and whether we keep investing in the Pi harness (Driver: Jonathan Mieloo)
  • Give people and agents a dependable local setup and an easy way to debug sandboxes and the task runtime (Driver: Jonathan Mieloo)
  • Track task and sandbox success rate and boot time on a dashboard, so we can show they're getting faster and more reliable (Driver: Jonathan Mieloo)
  • Investigate running the harness outside the sandbox and decide whether it's worth the move (Driver: Jonathan Mieloo)

Stretch

Internal evals (Drivers: Charles Vien Brandon Leung)

Motivation: Teams have no shared way to tell whether a model or agent change made things better or cheaper.

What we'll ship:

  • Give teams a shared way to test models and agent changes against real use cases, measuring quality and cost per successful task and catching regressions before they ship (Drivers: Charles Vien, Brandon Leung)

Ongoing work

These aren't objectives, but we're committing to them this quarter:

  • Make our services self-serve and more than one person deep.
  • Document the gateway, tasks, and sandbox interfaces.
  • Rewrite our runbooks and get the on-call rotation in good shape.
  • Add code owners for the tasks product folder and the agent package.

Handbook

We build the foundation that powers all of PostHog's LLM and agent applications. Product teams build agents and AI features on top of our infrastructure, so they can focus on their product instead of the plumbing underneath it.

What we own

  • AI gateway: the single endpoint PostHog's AI features use to call LLM providers, with shared auth and cost attribution
  • Tasks and sessions: the infrastructure that runs agents in cloud sandboxes
  • AI metering: real-time usage and spend tracking for every LLM call and agent run. We work with the Billing team to make sure that usage gets billed
  • Agent package: the shared agent runtime that other teams build their agents on
  • Hog harness: our custom agent harness, built on Pi. It captures product intent and passes it to the model, which leads to better generations. It ships as the Pi option in PostHog Desktop

Working with other teams

If a PostHog product uses an LLM or an agent, it probably runs on something we own.

Come to us if you:

  • Are calling an LLM from inside PostHog and want to go through the gateway
  • Are building an agent that needs a sandbox or a cloud session
  • Need metering or spend tracking for an AI feature, or think someone is abusing it
  • Hit a bug or a limit in the tasks infrastructure or the agent package

Slack channel

#team-agent-infra

Feature ownership

You can find out more about the features we own.