competitor_intel_agent

An autonomous AI agent that watches up to 5 competitor domains every morning — detecting new content, scoring it against your ICP, generating content ideas, and flagging pricing changes — then drops a structured report before your first coffee.

Built as part of my AI agent portfolio. Runs daily at 6am CST via Hermes (Nous Research's autonomous agent with a built-in cron scheduler). Companion to seo_agent_pipeline.

---

Results

Sample output from a live run (domains sanitized):

Domain	New Content	ICP Resonance	Ideas Generated	Pricing Change
rival-analytics.com	2 posts	High	3 ideas	None
adtracker-pro.com	0 posts	—	—	+$10/mo on Pro tier
creativespy.io	1 post	Low (filtered out)	0 ideas	None
marketlens.io	3 posts	High	5 ideas	None
spytools.co	0 posts	—	—	None

→ See the full sanitized report

---

What it does

config/targets.yaml
        │
        ▼
01 Sitemap Diff ──────────────────────► new_urls.json
        │
        ▼
02 Content Summarize ─────────────────► summaries.json
        │
        ▼
03 ICP Score ─────────────────────────► scored.json   (low-resonance filtered out)
        │
        ▼
04 Idea Generation ───────────────────► ideas.json

        ┌─ runs in parallel with 01–04 ─┐
        ▼                               │
05 Pricing Check ─────────────────────► pricing_diff.json

        │
        ▼
   Hermes Report ────────────────────► outputs/YYYY-MM-DD-report.md

Each stage writes a JSON artifact. The orchestrator reads the previous artifact before running the next stage — so a failed run resumes from the last completed step without re-doing work.

---

Design decisions

• Sitemap diff over crawling — sitemaps are machine-readable, fast, and reliable; full crawls are brittle and expensive at daily frequency
• ICP filter before idea generation — stage 03 discards low-resonance content before stage 04 runs; this cuts tokens by ~60% on noisy competitor blogs and keeps ideas on-target
• Stage files, not one giant prompt — each .md skill file has one job and bounded context; the orchestrator chains them and detects which stage to resume from on partial failure
• Snapshot files as memory — last_run/ snapshots replace a database; flat-file diffs are fast, auditable via git history, and require zero infrastructure
• Hermes as thin scheduler — the daily-run skill is a single-responsibility wrapper; changing cadence means one cron line, not code

---

Stack

• Hermes (Nous Research) — autonomous agent runtime with built-in cron scheduler; runs the pipeline daily at 6am CST
• OpenRouter — LLM routing; Hermes routes inference requests through OpenRouter
• Python 3.10+ — sitemap fetching (requests + xml.etree), HTML → plain text (markdownify), snapshot diffing
• YAML config — all domain targets, ICP definition, and output paths in one file; no hardcoding in stage files

---

File map

competitor_intel_agent/
├── README.md                          # This file
├── ARCHITECTURE.md                    # Design decisions in depth
├── config/
│   └── targets.example.yaml           # Domains, ICP definition, output paths
├── skills/
│   ├── competitor-intel.md            # Orchestrator skill (/competitor-intel)
│   └── stages/
│       ├── 01-sitemap-diff.md         # Fetch sitemaps → diff → new_urls.json
│       ├── 02-content-summarize.md    # Fetch new URLs → summaries.json
│       ├── 03-icp-score.md            # Score against ICP → scored.json
│       ├── 04-idea-generation.md      # Generate content ideas → ideas.json
│       └── 05-pricing-check.md        # Scrape pricing page → pricing_diff.json
├── hermes/
│   └── daily-run.md                   # Scheduler wrapper (6am CST cron)
├── last_run/                          # Snapshot memory (gitignored except .gitkeep)
└── outputs/
    └── example-report-2026-05-09.md   # Sanitized sample run output

---

Hermes scheduling

Hermes is an autonomous AI agent by Nous Research with a built-in cron scheduler that ticks every 60 seconds and executes due jobs in isolated agent sessions.

The hermes/daily-run.md file defines the cron job: what to clear, what to run, and how to deliver the report. Scheduling concerns live in Hermes; pipeline logic lives in the stage files. Swapping from daily to twice-daily means changing the schedule in Hermes, not touching any pipeline code.

---

How to adapt it

1. Fork the repo
2. cp config/targets.example.yaml config/targets.yaml — fill in your domains and ICP definition
3. Create a Hermes cron job pointing at hermes/daily-run.md, scheduled for 6am your timezone
4. Optionally adjust ICP scoring thresholds in skills/stages/03-icp-score.md

The agent becomes significantly more useful with additional context: your existing content inventory, brand voice, and GSC performance data can all feed into the ICP scoring and idea generation stages.

---

Requirements

• Hermes with an OpenRouter API key
• Python 3.10+
• pip install requests markdownify pyyaml

Interested in doing this for your business or looking to add powerful AI GTM flows? Email me at loganriebel@gmail.com or connect with me at https://www.linkedin.com/in/logan-riebel/