Pineapple on pizza

No data available

Members

Roadmap

Recently shipped

Trace Firebase Genkit flows

Firebase Genkit flows can send model, tool, and embedding traces to AI observability. The integration records tokens, costs, messages, and errors while redacting credentials.

Goals

Q4 2026 objectives

Objective 1: Finish the UI redesign Carlos Marchal

Motivation: Every tab in the core tracing product should help users find issues with their AI agents and features. The redesign will make these tabs more useful and provide a consistent experience across the product.

What we'll ship:

  • UI redesign – redesign all tabs in the core tracing product around finding issues with AI agents and features, with consistent navigation and interactions
  • Matching MCP tools and skills – add tools and skills that mirror the new UI pages, so coding agents can find the same high-priority issues with AI agents and features through MCP

Objective 2: Roll out offline evals Radu Raicea

Motivation: Offline evals shipped internally in Q3. We now need to roll them out to external users, iterate on their feedback, and reach general availability.

What we'll ship:

  • External rollout and iteration – roll out offline evals to external users and improve the experience based on their feedback
  • Billing and data retention – add billing support and apply data retention policies to offline evaluation data
  • General availability – bring offline evals to general availability after the external rollout and iteration

Objective 3: Multimodal general availability Carlos Marchal

Motivation: Multimodal ingestion and trace views are available. We need to make them generally available with clear limits and pricing, and support large text content too.

What we'll ship:

  • General availability – finalize the cap and pricing for multimodal content
  • Large text content – extend multimodal storage and ingestion support to large text inputs and outputs

Objective 4: Query and loading optimization

Motivation: Slow queries and loading times make investigations frustrating. We need improvements in both the UI and the underlying queries.

What we'll ship:

Objective 5: Useful clusters Radu Raicea

Motivation: Clusters need to surface useful, actionable findings, and their cost needs to make sense for the product.

What we'll ship:

  • Opt-in clusters – make clustering opt-in
  • Customizable summarization prompts – give teams control over how their traces are summarized for clustering
  • Richer cluster centroids – enrich centroids with evaluation results and show how they evolve over time
  • Pricing exploration – explore pricing that makes clustering sustainable

Objective 6: Online evals improvements Bernat Torres

Motivation: Teams need evaluations that go beyond boolean results, work with multimodal content, and make the results easy to understand.

What we'll ship:

  • More output types – support categorical, multi-categorical, and numerical evaluation results
  • Multimodal evaluations – evaluate multimodal content
  • Results reporting – improve the UI for reporting evaluation results

Objective 7: Use case / agent primitive Bernat Torres

Motivation: Teams need to distinguish between the different use cases and agents in their product to understand their behavior.

What we'll ship:

  • Use case / agent primitive – introduce a way to organize AI Observability data by use case or agent

Handbook

Who we're building for

Product Engineers and Full Stack Developers who are building:

  • AI-native products (agents, assistants, copilots, specialized hardware)
  • AI-adjacent products (LLMs integrated into existing products)

AI Observability is a good fit if:

  • They need to monitor traces, spans, token costs, latency, and analyze usage of AI features
  • They're using trace summaries to debug and evals to make product decisions
  • They care about questions like: “how does interacting with LLM features correlate with retention, usage, or revenue?”
  • They're already PostHog users (or should be) using product analytics and session replay to combine qualitative context with quantitative data
  • They want to start getting value right away, without needing extensive setup and configuration

Who else might want to use AI Observability at their org:

  • Application Ops / SRE to monitor production AI systems for errors, prompt injection, jailbreaks, or other security issues
  • Product Managers to understand user sentiment, usage and make decisions about their AI roadmap
  • Customer Success / Support Teams to improve documentation or investigate user issues

Who we're NOT building for (right now)

AI Researchers and Machine Learning (ML) Engineers doing:

  • Deep foundation model work
  • Complex benchmarking and evaluation
  • Advanced experimentation requiring specialized tooling

These folks are running CI/CD pipelines and building full QA automation frameworks with performance benchmarks. If they try us and churn, that's fine. We haven't built the tools they need (yet).