PM Brief
🔍 Click image to zoom in
The Problem
Product managers start every day the same way — opening 8-10 tools to figure out what happened overnight. Email, calendar, Jira, Slack, competitive blogs, analytics dashboards. The data exists across all of them, but the triage is manual, repetitive, and fragmented.
What if AI agents could do the overnight triage autonomously — scanning every source, filtering signal from noise, cross-correlating across systems, and delivering a single prioritized briefing by the time the PM opens their laptop?
Not a dashboard that displays data. An intelligence layer that thinks about it.
This project demonstrates how multi-agent orchestration enables autonomous information gathering and synthesis — the same pattern that would power internal decision-support tools at any product-led company.
Architecture
A hierarchical supervisor pattern: one Synthesizer agent (Claude Sonnet) coordinates four specialized worker agents (GPT-4.1 nano), each connected to a different data source. The workers handle high-token ingestion on cheap models; the Synthesizer handles cross-source reasoning on a strong model.
Gmail via n8n native node
Scans inbox, classifies messages by urgency, filters noise (newsletters, FYIs, cc'd threads), surfaces action items.
Google Calendar via n8n native node
Detects schedule changes, cancelled meetings, and time conflicts. Recommends keep/skip/prep for each event.
Jira via n8n native node
Monitors ticket queue for escalations, P1 bugs, and status changes. Flags items requiring PM attention.
Tavily web search via HTTP Request
Runs targeted searches for competitor product updates, pricing changes, and announcements.
Claude Sonnet via Anthropic API
Receives condensed outputs from all four workers. Cross-references signals, ranks by priority, identifies correlations across sources, and produces a structured briefing with recommended actions.
The workflow runs on n8n, self-hosted via Docker on a Mac Studio. Sequential execution — each agent runs in order, results flow up to the Synthesizer. No external orchestration framework, no cloud deployment. The entire system runs locally.
How It Works
Agents Scan Overnight
The workflow triggers and each agent independently connects to its data source — Gmail, Google Calendar, Jira, and Tavily. They ingest raw data using GPT-4.1 nano, classify and filter it, then pass condensed summaries to the Synthesizer.
Synthesizer Cross-References
Claude Sonnet receives all four agent reports and identifies connections between them — a Jira P1 ticket that matches an escalation email, a calendar meeting that needs prep based on competitive intel. Signals from one source inform the priority of another.
Prioritized Briefing Generated
The Synthesizer produces structured JSON — ranked priority items with recommended actions, calendar events with keep/skip/prep tags, an overnight changelog, and summary statistics. "3 need action" not "47 emails."
PM Reviews One Screen
The briefing populates a dashboard where the PM can scan everything in under 2 minutes — compared to 60+ minutes manually triaging across 8-10 tools. Each item links back to its source for drill-down.
Before & After
To demonstrate the impact of the multi-agent system, we designed a custom PM dashboard in Figma that visualizes exactly what the agents produce. The PM checked in Wednesday at 6:00 PM — nothing new. Overnight, four agents worked autonomously. By Thursday 8:15 AM, the dashboard shows what changed and what needs attention.
🔍 Click image to zoom in
Wednesday 6:00 PM — Last Check-In
What Each Agent Found
Each annotated screenshot isolates one agent's contribution — showing how a single detection ripples across the entire dashboard. Numbered markers connect to the summary box.
🔍 Click any image to zoom in
Calendar Agent — 2 Changes Detected
Email Agent — 47 Scanned, 4 Surfaced
Support Agent — 1 P1 Escalation
Competitive Intel Agent — 2 Releases Detected
What the System Revealed
Beyond the individual agent reports, the Synthesizer surfaced patterns that would take a PM significant time to piece together manually.
Cross-source correlation catches what siloed tools miss
The APAC checkout error appeared in both email (from James Liu at 2:14 AM) and Jira (P1 ticket). The Synthesizer confirmed it was the same incident from two independent sources and escalated it to the top of the priority list. No single tool tells that story alone.
Filtering is the real value — not aggregation
The Email Agent scanned 47 messages and surfaced 4. 43 were noise — newsletters, FYIs, cc'd threads. The intelligence isn't in showing everything; it's in knowing what to hide. "3 need action" is more useful than "47 unread."
Context-aware recommendations beat status updates
The system didn't just detect that VP Strategy Sync moved to 3:00 PM — it connected that meeting to the Competitive Intel findings and recommended the PM prep with the competitor pricing data. The calendar change informed the priority list, which informed the meeting prep. One change, traced across four sections.
Limitations & Production Considerations
Test data, not live production systems
This demo uses seeded test data across Gmail, Google Calendar, and Jira. In production, the agents would connect to the same APIs with real data. The architecture is identical — only the volume and variability of data changes.
Sequential execution, not parallel
Agents run one after another for easier debugging and deterministic behavior. For a daily batch workflow, sequential is sufficient. Parallel execution would reduce total runtime but adds coordination complexity — a worthwhile optimization for real-time use cases.
Dashboard is a Figma mockup
The frontend is a design artifact, not a working application. The value proposition is the intelligence layer — the multi-agent orchestration, cross-source reasoning, and structured output. The dashboard is a commodity; building it in React would demonstrate frontend skills, not AI product thinking.
Known failure modes
The Email Agent may over-prioritize messages with urgent keywords without sufficient context. The Competitive Intel Agent may flag irrelevant posts due to keyword similarity. The Synthesizer may overweight certain sources depending on prompt framing. Mitigations include structured output schemas, confidence scoring, and threshold filtering.
Vision — What Comes Next
The current system demonstrates the core pattern with 4 agents. In production, this architecture extends to any data source a PM touches.