Agentic PM Kickoff | PostHog × Linear × Supabase × Dreambase

Agentic feature discovery

How agents are powering product discovery, not just delivery

Marco G

PM, AI Observability · PostHog

What is PostHog?

Your product data, in one place

Get insights across your data sources.

Product analytics Web analytics Session replay Feature flags Experiments Surveys Error tracking Logs AI observability Data warehouse Support Metrics Workflows + more

The shift

Self-driving = discovery + delivery

Agents are now working on the discovery side of the product loop.

Discovery agents

Signals and scouts

Reactive

Signals

Something happened: a pattern in the data becomes a signal you can act on.

Online Evaluations · error spikes · survey drops

Proactive

Scouts

Nothing's broken yet: scouts go looking for the users and behavior worth a closer look.

costly or unusual users

The example · a real customer, anonymized

Excel for AI

An agent that assists with spreadsheet creation and formula generation.

Feature

Spreadsheet creation

Draft a full sheet from a plain-English prompt.

Feature

Formula generation

Formulas written and fixed on demand.

The example

Known features, unknown wants

You understand what you've built. The full set of things users want is fuzzy.

The pipeline

From traces to a report

Traces split into clusters and sentiment, then a scout joins them into a report.

Step 1 · Traces

Messages show your users' wants

You can see intent and sentiment.

User"Can you help me reconcile these two tabs?"

Assistant"Sorry, I can only help with structure, formatting, and formulas."

User"Ugh. That's what I actually needed."

→

Intent help with accounting work

Sentiment negative

Step 2 · Clusters

Similar messages become clusters

Grouped together by embeddings.

Ask reconcile these ledgers

Ask book my journal entries

Ask help with month-end close

Ask build a DCF model

+ thousands more conversations

→

Cluster

Help with accounting work

418 conversations

Cluster

Financial modeling

5.2k conversations, the expected use

Step 3 · Sentiment evals

Evals grade message sentiment

With a score of positive, neutral, or negative.

✕ "NO THAT'S WRONG"

~ "ok, moving on"

✓ "perfect, thanks!"

→

✕ Negative

Help with accounting

418 conversations

✓ Positive

Financial modeling

5.2k conversations

Step 4 · Scouts

A scout investigates and writes a report

Scout · runs nightly

"Monitor clusters (especially new ones with negative sentiment) to see if there's any surface areas we don't support well."

→

Discovery scout APP 9:03 AM

🔎 Found something: help with accounting

418 conversations with negative sentiment are asking for accounting workflows the product doesn't support.

View the full report →

Takeaways

What to take home

Agents are getting good at discovery, making self-driving possible.
Engineers supervise delivery agents; PMs supervise discovery agents.
PMs create agentic discovery by building the high-quality context pipelines: evals, signals, scouts.

Thank youQuestions?

AI observability · posthog.com/ai-observability
Slides · posthog.github.io/presentations/agentic-feature-discovery

Find me

marco.gancitano.com
github.com/mgancita