🤖 PM Research Agent

An agentic AI workflow that autonomously researches a company or product and generates a structured competitive brief — the kind a PM would spend 3–4 hours assembling manually, done in under 2 minutes.

---

Why I built this

Competitive research is one of the highest-leverage activities a PM can do — and one of the most time-consuming to do well. A proper competitive brief requires:

• Tracking down recent pricing changes
• Reading through product announcements and release notes
• Synthesizing customer reviews and analyst takes
• Framing it all in terms of strategic implications, not just feature lists

I built this agent to automate the research and synthesis layer, so I can spend my time on the part that actually requires product judgment: deciding what it means for roadmap, positioning, and customer conversations.

This is also a working prototype of a broader pattern I care about: agentic workflows that augment PM decision-making, not just generate text.

---

What's in this repo

Module	What it does
agent.py	Core research agent — autonomously searches and generates a competitive brief
scorer.py	Scores a brief across 5 strategic dimensions and outputs a threat level
batch.py	Researches and scores multiple competitors in one run, outputs a ranked landscape
diff.py	Compares two briefs over time, surfaces strategic changes and threat trajectory
outputs/slack_output.py	Pushes briefs and scorecards to a Slack channel via webhook
outputs/notion_output.py	Creates a Notion database page for each brief with scorecard properties

---

How it works

1. Research Agent (agent.py)

Runs an autonomous multi-step research loop using Claude + web search:

User Input (target company/product)
        │
        ▼
┌─────────────────────────────────┐
│        Research Agent Loop      │
│  1. Decide what to search for   │
│  2. Run web search              │
│  3. Evaluate results            │
│  4. Decide: search more or stop │
│  5. Repeat up to 10x            │
└─────────────────────────────────┘
        │
        ▼
Structured Competitive Brief (Markdown)

The agent decides its own search strategy. It doesn't follow a fixed script — it reads intermediate results and determines what gaps remain before synthesizing.

2. Scoring Layer (scorer.py)

Takes a brief and scores the competitor across five strategic dimensions:

Dimension	What it measures
Market Overlap	How directly do they compete for the same customers and budget?
AI Maturity	How deeply is AI integrated — native or bolted on?
Execution Velocity	How fast are they shipping meaningful product changes?
Distribution Strength	Brand, channels, partnerships, existing contracts
Resource Depth	Funding, headcount — how long can they sustain a fight?

Output: a composite score (1–10) + overall threat level: 🔴 Critical / 🟠 High / 🟡 Medium / 🟢 Low

3. Batch Mode (batch.py)

Research and score an entire competitive landscape in one command. Outputs:

• Individual brief files per competitor
• landscape.md — a ranked summary table sorted by threat level

4. Diff Mode (diff.py)

Compare two briefs for the same competitor across time. Surfaces:

• Threat trajectory (↑ Escalating / → Stable / ↓ De-escalating)
• New features and strategic moves
• ICP drift — are they moving toward your customers?
• AI maturity delta
• Strategic implications for a competing PM

5. Outputs (Slack + Notion)

Push briefs and scorecards directly to your workspace:

• Slack: Formatted Block Kit message to any channel via webhook
• Notion: New database page with scorecard properties, filterable by threat level

---

Quickstart

Install

git clone https://github.com/dvmukul/pm-research-agent.git
cd pm-research-agent
pip install -r requirements.txt
export ANTHROPIC_API_KEY=your_key_here

Research a single competitor

python agent.py --target "Salesforce Einstein"
python agent.py --target "Intercom Fin AI" --output briefs/intercom.md

Score a brief

python scorer.py --brief briefs/intercom.md
python scorer.py --brief briefs/intercom.md --json   # structured JSON output

Research + score in one step

python agent.py --target "Notion AI" --output briefs/notion.md && \
python scorer.py --brief briefs/notion.md

Run a full competitive landscape

# Comma-separated
python batch.py --targets "Notion AI, Coda, Confluence AI, ClickUp AI" --output-dir briefs/

# From a file
python batch.py --targets-file competitors.txt --output-dir briefs/ --context "Q2 2025 review"

Diff two briefs over time

python diff.py --before briefs/notion-jan.md --after briefs/notion-apr.md --target "Notion AI"
python diff.py --before briefs/notion-jan.md --after briefs/notion-apr.md --output diffs/notion-delta.md

Push to Slack

export SLACK_WEBHOOK_URL=https://hooks.slack.com/...
python outputs/slack_output.py --brief briefs/notion.md --target "Notion AI"
python outputs/slack_output.py --brief briefs/notion.md --dry-run   # preview without sending

Push to Notion

export NOTION_API_KEY=secret_xxx
export NOTION_DATABASE_ID=your-database-id
python outputs/notion_output.py --brief briefs/notion.md --target "Notion AI" --category "AI"

---

Example outputs

Terminal output (batch run)

📋 Batch run: 3 competitor(s)
   Targets: Notion AI, Coda, Confluence AI
   Output:  briefs/

════════════════════════════════════════════════════════════
  [1/3] Notion AI
════════════════════════════════════════════════════════════
  🌐 Search [1]: Notion AI features pricing 2025
  🌐 Search [2]: Notion AI vs Microsoft Copilot
  🌐 Search [3]: Notion enterprise announcements 2025
  🎯 Scoring threat level for: Notion AI
     🟠 High threat  |  Composite: 7.4/10

════════════════════════════════════════════════════════════
  [2/3] Coda
...

✅ Batch complete
   3 brief(s) saved to briefs/
   Landscape summary: briefs/landscape.md

Scorecard output

## Threat Scorecard: Notion AI

**Overall Threat Level:** 🟠 High
**Composite Score:** 7.4 / 10

| Dimension          | Score  | Visual       | Rationale                                          |
|--------------------|--------|--------------|----------------------------------------------------|
| Market Overlap     | 8/10   | ████████░░   | Direct overlap on knowledge management and docs    |
| AI Maturity        | 7/10   | ███████░░░   | Native AI on structured data is a real moat        |
| Execution Velocity | 8/10   | ████████░░   | Notion Mail + Calendar in one quarter signals pace |
| Distribution       | 7/10   | ███████░░░   | Strong SMB brand but enterprise flank exposed      |
| Resource Depth     | 7/10   | ███████░░░   | $10B valuation, well-capitalized                   |

Full example brief: sample_output/notion-ai-brief.md

---

Design decisions worth noting

Why Claude? Claude's extended context window and strong instruction-following made it the right choice for a task that requires reading multiple search results and synthesizing them coherently — not just summarizing the top result.

Why agentic (not a fixed pipeline)? Fixed pipelines break when the target is unusual or when the first search returns stale results. The agent adapts — if it can't find pricing, it searches differently. If a product was acquired recently, it notices and adjusts the framing.

Why five scoring dimensions? They were chosen to map directly to the questions a PM asks in a real competitive review — not generic "strengths/weaknesses." Each dimension is specific enough to score from a brief, and together they produce a threat profile, not just a threat number.

Why a diff mode? A one-time brief is a snapshot. What matters to a PM is momentum — is this competitor accelerating or stalling? The diff mode is what turns this from a research tool into an ongoing intelligence system.

Why Markdown output? Briefs need to be shareable. Markdown renders cleanly in Notion, Confluence, GitHub, and Linear — everywhere a PM actually works.

---

About

Built by Mukul Dewangan — Senior PM specializing in AI and data products. This repo is part of a broader exploration of how agentic AI can augment product management workflows.

If you're building in this space, let's connect.