DraftThis concept page is in progress. Discovery Agent and the knowledge graph are shipped; the polished cross-product upload-to-backlog flow shown here is the next refinement.
Connected Artifacts, Designed
The unified knowledge graph is the bet underneath everything in ProductIntel. Stories, docs, cases, requests, APIs, learned heuristics — same artifact type, same edges, same queryable graph. The point of the bet is not the data model. It is what becomes possible when AI can reason across all of it at once.
What follows is one designed example of what that reasoning looks like. A single document in. Three product-aware recommendations out. Five minutes from upload to a queued backlog of work units that already know which files to touch in which product.
The Scenario
One PM manages three products built on the same ProductIntel instance: Easycomm, Acme CRM, and Acme Insurance. They drop a single document into the platform — the Postmark v4 API release notes — and let the Discovery Agent loose on it. What follows is what happens when every artifact in the system already lives in the same graph.
ProductIntel · Discovery Agent · Intake
Drop a doc to discover
API specs, changelogs, RFCs, design notes. The agent decides what to do with it.
postmark-v4-release-notes.md
38 KB · Markdown · API release notes
✓ Ready to process
Discovery action recognized
This looks like an API release notes doc. Recommended action: scan all enabled products for upgrade and opportunity matches. The agent will read your codebase against the changes and surface per-product recommendations.
What humans do
Drags the API doc into the Discovery Agent intake. Confirms scope (which products to consider) — defaults to all enabled products.
What AI does
Recognizes the document type (API spec / changelog / design doc), surfaces an appropriate Discovery action, requests confirmation before doing anything irreversible.
ProductIntel · Discovery Agent · Reading
Discovery activity
⬤ RUNNING · 23s elapsed
What humans do
Watches a short activity log and waits. There is nothing to decide here yet.
What AI does
Parses the API spec into a structured shape (endpoints, methods, breaking changes). Queries the knowledge graph for any artifact that mentions related concepts. Cross-references against existing code in each product.
ProductIntel · Discovery · Connections Detected
Knowledge graph · new edges materialized
12 edges · 3 typed relationships
Direct upgrade
2 files · Postmark v3 → v4 SDK
Upgrade + feature
3 files · v3 → v4 + new sender authentication API
Opportunity
0 files · case digest could use v4 batching for high-volume sends
→ Direct upgrade — files already use v3, mechanical migration
→ Upgrade + feature — existing usage maps to new capability
→ Opportunity — net-new applicability, no existing dependency
Why three different edge types matter
A flat “here are matches” list collapses very different kinds of work into one bucket. The graph distinguishes them so the recommendation phase can shape each appropriately. Direct upgrades become small mechanical stories. Opportunity gaps need product judgment before they become work.
What humans do
Reviews the detected connections. Can prune anything that looks wrong before recommendations are generated.
What AI does
Materializes typed edges in the knowledge graph between the new API doc and existing artifacts. Different edge types: direct dependency, pattern match, opportunity gap.
ProductIntel · Discovery · Recommendations
Per-product recommendations · 3
All grounded in existing code
Migrate digest emails from Postmark v3 SDK to v4
Mechanical SDK upgrade. v4 collapses two send calls into one and adds the new template engine. No behavior change for end users.
Files anchored
Size · Verdict
Small · ~14 LOC delta
Agent team can take this
Migrate to Postmark v4 + adopt new sender authentication API
Direct SDK upgrade plus a real feature opportunity: v4 sender auth eliminates the per-domain DKIM verification flow your sales team currently does manually for new clients.
Files anchored
Size · Verdict
Medium · ~80 LOC + admin UI changes
Hybrid: agents take SDK upgrade, humans take admin UI
Replace Resend with Postmark v4 for case digest cron (opportunity)
Acme Insurance does not use Postmark today. The case digest cron is on Resend, which has been throttling at month-end. Postmark v4 batching would handle the load cleanly. This is an opportunity, not a forced migration.
Files anchored
Size · Verdict
Medium · greenfield service + cron rewrite
Needs PM judgment · effort vs Resend pain
What humans do
Reads the three recommendations side by side. Adjusts scope, accepts, or sets aside any of them.
What AI does
Synthesizes per-product recommendations from the connected edges. Each recommendation includes the type of work, the existing code it touches, the rough size, and the why.
ProductIntel · Work · Backlog · Filtered to Discovery-sourced
New work units · pushed from Discovery
✓ 3 of 3 created · linked to source doc
Migrate digest emails from Postmark v3 SDK to v4
Migrate to Postmark v4 + adopt new sender authentication API
Evaluate Postmark v4 for case digest (Resend replacement)
Lineage preserved
Each work unit links back to the source doc and to the existing files it touches. When an agent or a human picks one up, the spec, the API doc, and the relevant code patterns are already in context. No re-discovery needed.
What humans do
Clicks ‘Push all to backlogs’. Reviews the result in the unified backlog view. Tags or schedules as needed.
What AI does
Creates the work units against each product’s backlog with team tags, file anchors, and the original doc as a parent reference. Each work unit is already refined enough to be picked up by an agent or a human in the next sprint.
Five Minutes, Three Products
In a federated stack, this story does not happen. The API doc lives in Confluence, the existing integrations live in code that nobody indexed for AI, the products live in different Jira projects, and the cross-product reasoning has to happen in someone’s head over the course of a week. Most weeks, it does not happen at all.
The unified graph is a small architectural choice that pays off in surprising places. This is one of them. The Knowledge Chat answering a question that spans three products is another. The Discovery Agent finding a backlog cluster from a customer case is another. Same graph, different expressions.