Okara vs Crystal-Agents: Strategic Enhancement Analysis

Executive Summary

Okara built a vertical AI CMO product — a domain-specific multi-agent system dedicated entirely to marketing automation for SaaS founders. Crystal-Agents is a horizontal agent stack — a general-purpose workflow engine that can handle coding, marketing, ops, and sales.

The architectural patterns are nearly identical. The difference is in packaging, depth, and autonomy. Okara proves that Crystal-Agents’ existing architecture is the right foundation — but also reveals six concrete areas where we can deepen our capabilities significantly.

---

What Okara Does (Broken Down)

Okara Component	What It Does	Automation Level
CMO Agent (root)	Reads your website, maps competitors, learns brand voice, drafts strategy	Runs on signup — fully autonomous initial audit
SEO Agent	Technical site audit, keyword classification (TOF/MOF/BOF), fix tracking	Audit is autonomous; fixes are surfaced to user
GEO Agent	Tracks citations in ChatGPT, Claude, Perplexity, Grok	Fully autonomous monitoring
Coding Agent	Reads codebase via GitHub, writes PRs to fix SEO/GEO issues	User clicks “Fix it” → agent creates PR → user merges
Article/Writer Agent	Drafts SEO+GEO-optimized articles in brand voice	Drafts autonomously, publishes on CMS connect
UGC Agent	Generates video scripts from product context	Autonomous drafting
Reddit Agent	Monitors subreddits, scores threads, drafts replies	Surfaces 2 best opportunities/day with draft replies
LinkedIn Agent	Drafts posts in founder’s voice about real product journey	Drafts queued in feed; user reviews + posts
X Agent	Reads trending niche content, drafts posts grounded in product	Same as LinkedIn — daily draft queue
llms.txt Generator	Free tool — generates llms.txt file for any website	One-shot utility

---

Architecture Comparison: What We Already Have

graph LR
    subgraph "Okara Architecture"
        O_CMO["CMO Agent
(Orchestrator)"]
        O_SEO["SEO Agent"]
        O_GEO["GEO Agent"]
        O_CODE["Coding Agent"]
        O_WRITER["Writer Agent"]
        O_UGC["UGC Agent"]
        O_REDDIT["Reddit Agent"]
        O_LI["LinkedIn Agent"]
        O_X["X Agent"]
        O_CMO --> O_SEO & O_GEO & O_CODE & O_WRITER & O_UGC & O_REDDIT & O_LI & O_X
    end

    subgraph "Crystal-Agents Architecture"
        C_CEO["CEO Router
(Orchestrator)"]
        C_DEV["Developer"]
        C_DEVOPS["DevOps"]
        C_CONTENT["Content"]
        C_RESEARCH["Research"]
        C_SEO["SEO"]
        C_ADS["Ads Manager"]
        C_SALES["Sales"]
        C_TESTER["Tester"]
        C_REVIEWER["Reviewer"]
        C_ANALYTICS["Analytics"]
        C_CEO --> C_DEV & C_DEVOPS & C_CONTENT & C_RESEARCH & C_SEO & C_ADS & C_SALES & C_TESTER & C_REVIEWER & C_ANALYTICS
    end

Direct Capability Mapping

Okara Agent	Crystal-Agents Equivalent	Current Status
CMO Agent (orchestrator)	ceo_router.py	✅ Built — routes NL to workflows
Coding Agent	developer.py + git_ops.py	✅ Built — branch, commit, push, PR
Writer Agent	content.py	⚠️ Exists but generic — no SEO/GEO optimization
SEO Agent	seo.py	⚠️ Stub — no audit, no keyword classification
GEO Agent	(not present)	❌ Missing entirely
UGC Agent	(not present)	❌ Missing
Reddit Agent	(not present)	❌ Missing
LinkedIn Agent	(not present)	❌ Missing
X Agent	(not present)	❌ Missing
CMS Integration	(not present)	❌ No CMS publish tool
GA/GSC Integration	(not present)	❌ No analytics data ingestion
Action Feed (dashboard)	app.py	⚠️ Shows runs/approvals, but no “action feed” pattern

---

What Okara Does That We Don’t — And Should

1. Autonomous Onboarding Audit

When a user signs up for Okara, the CMO agent immediately:

• Reads the website end-to-end
• Maps the competitive landscape
• Learns the brand voice
• Drafts a marketing strategy
• Runs a technical site audit

Crystal-Agents equivalent: We have memory/ with static .md files (company, product, pricing, soul, marketing). But these are manually written. There’s no autonomous “learn the business” workflow.

Recommendation: Build a business_onboarding workflow that:

1. Takes a website URL as input
2. Uses web_search to crawl and extract product info, pricing, competitors
3. Auto-populates memory/company.md, memory/product.md, memory/pricing.md
4. Generates a draft memory/marketing.md strategy
5. Outputs a “Business Profile” report to generated/

This turns Crystal-Agents from “you configure me” to “I configure myself.”

---

2. Action Feed Pattern (Not Just Run History)

Okara’s dashboard isn’t just a log viewer. It’s an Action Feed — a prioritized list of things the user should do, each with a one-click action button.

Example: “Your site is missing meta descriptions on 12 pages. [Fix it]”

Our dashboard (app.py) shows runs, approvals, and scheduler history. That’s operational telemetry, not a decision-making surface.

Recommendation: Add an Action Feed concept:

• A new ActionItem dataclass: {type, title, description, severity, suggested_workflow, suggested_prompt, status}
• Agents write action items as they discover opportunities (SEO issues, content gaps, social opportunities)
• Dashboard renders them as a prioritized feed with “Run” buttons
• Each “Run” pre-fills the workflow form with the right workflow + prompt

This is the key difference between a tool the user has to drive vs. a tool that drives the user.

---

3. Platform-Specific Social Agents

Okara has dedicated agents for Reddit, LinkedIn, and X, each with distinct behaviors:

Platform	Agent Behavior
Reddit	Monitor subreddits → score threads → draft contextual replies
LinkedIn	Read founder activity → draft thought-leadership posts in their voice
X	Read niche trends → draft posts grounded in product + POV

Crystal-Agents has a generic content.py agent. The marketing.md memory file mentions LinkedIn/Social/Blog but the agent doesn’t differentiate.

Recommendation: Don’t build 3 separate agent classes. Instead:

• Keep the single content agent but add platform-specific prompt templates in memory/ or workflows/templates.py
• Build a social_content workflow with steps: research → draft → review
• Add a platform parameter to the workflow: linkedin | x | reddit | blog
• Each platform template encodes the best practices (LinkedIn = hook in first 2 lines; Reddit = authentic reply tone; X = concise + punch)

This keeps our architecture lean while delivering Okara’s per-platform specialization.

---

4. SEO / GEO Agent Intelligence

This is Okara’s deepest capability and our biggest gap.

What Okara’s SEO Agent does:

• Full technical site audit (meta tags, structure, performance)
• Keyword classification by funnel stage (Top/Middle/Bottom of Funnel)
• Issue tracking: which fixes moved the needle
• Integration with Google Analytics and Google Search Console

What Okara’s GEO Agent does (entirely novel):

• Tracks whether your product is cited by ChatGPT, Claude, Perplexity, Grok
• Monitors 10+ prompts to see if AI models recommend you
• This is a new marketing channel — “Generative Engine Optimization”

Crystal-Agents today: seo.py is a stub inheriting from BaseAgent. Our web_search tool exists but isn’t wired for site auditing.

Recommendation — phased approach:

Phase A (achievable now): Build a site_audit workflow:

1. Use web_search to crawl the target site’s key pages
2. LLM analyzes: missing meta descriptions, heading structure, missing alt text, page speed indicators, missing llms.txt
3. Outputs a report to generated/seo_audit.md
4. Creates ActionItem entries for each issue found

Phase B (needs API keys): Add real GA/GSC integration:

• New tool: tools/analytics.py wrapping the Google Analytics Data API
• Ingest traffic data, keyword rankings, click-through rates
• Feed this into the SEO agent’s context for data-driven recommendations

Phase C (advanced): Build the GEO Agent concept:

• New tool: tools/geo_monitor.py
• Query multiple LLM APIs with buying-intent prompts (“best ERP for manufacturing”)
• Parse whether your product appears in responses
• Track changes over time in memory/history/geo_citations/

---

5. Content → Publish Pipeline

Okara doesn’t just draft content — it publishes. Connect your CMS and articles auto-publish.

Crystal-Agents drafts content to generated/. That’s where it stops. The human has to manually copy-paste to their CMS.

Recommendation: Build a publish tool abstraction:

• tools/publisher.py with adapters for common targets:
  • wordpress (REST API)
  • ghost (Admin API)
  • file (write to a deployment directory)
  • markdown_repo (commit to a docs repo via git_ops)
• The content_drafting workflow gets a new final step: “publish” (with requires_approval: true)
• This closes the loop: draft → review → approve → live

---

6. Coding Agent as a Marketing Tool

This is Okara’s cleverest insight: the coding agent doesn’t serve engineering — it serves marketing.

Their coding agent exists specifically to fix SEO/GEO issues discovered by the audit agents. It reads the codebase, writes the fix, opens a PR. The loop is: audit finds problem → coding agent fixes it → user merges.

Crystal-Agents already has this capability in Phase 2 (git_workflow.py). What we’re missing is the trigger loop:

Recommendation: Wire the site_audit workflow’s output (ActionItems) to the git_workflow as an automatic follow-up:

• When an SEO issue is discovered that maps to a code fix (missing meta tag, broken heading hierarchy, missing robots.txt)
• The Action Feed shows a “Fix it” button
• Clicking it triggers git_workflow with a pre-filled prompt derived from the audit finding
• This creates the same audit → fix → PR → merge loop that Okara has

---

What We Should NOT Copy

Okara Feature	Why We Skip It
UGC video scripts	Highly niche SaaS-marketing feature. Low priority for a general agent stack.
Multi-seat team plans	Product/pricing concern, not architecture. Revisit if/when we monetize.
“Growth accelerator” program	Community/business feature, not technical.
Real-time subreddit monitoring	Requires persistent polling + Reddit API credentials. High infra cost for uncertain ROI. Consider only after social agents prove value via manual triggers.

---

Recommended Phased Roadmap

Phase 3A — Foundation (builds on Phase 2)

#	Enhancement	Key Files	Effort
1	Action Feed data model + dashboard rendering	orchestrator/action_feed.py, dashboard/app.py	Medium
2	Business Onboarding workflow	workflows/business_onboarding.py, memory/ auto-populate	Medium
3	Site Audit workflow (SEO basics)	workflows/site_audit.py, agents/seo.py filled in	Medium
4	Platform-specific content templates	memory/templates/linkedin.md, reddit.md, x.md, blog.md	Small

Phase 3B — Integration

#	Enhancement	Key Files	Effort
5	Publisher tool (WordPress/Ghost/file adapters)	tools/publisher.py	Medium
6	Audit → Fix → PR loop (wire Action Feed to git_workflow)	dashboard/app.py, workflows/git_workflow.py	Small
7	GA/GSC integration tool	tools/analytics.py	Large

Phase 3C — Advanced

#	Enhancement	Key Files	Effort
8	GEO Monitor (track AI model citations)	tools/geo_monitor.py, agents/geo.py	Large
9	Social scheduling (queue drafts for timed posting)	scheduler/, tools/social_poster.py	Large
10	Consolidated daily brief (morning report with all signals)	reports/summary.py expanded	Medium

---

Architecture Principles to Maintain

Our system has structural advantages over Okara that we should protect:

1. Horizontal, not vertical: Okara is locked to marketing. Crystal-Agents handles dev + ops + marketing + sales. Keep the general-purpose agent/tool/workflow pattern.

2. HITL by default: Okara’s “Fix it” button is appealing, but our requires_approval + ApprovalStore pattern is more robust. Every new tool should respect the policy.json gate.

3. Memory as markdown files: Okara likely uses a database. Our memory/ approach (company.md, soul.md, etc.) is simpler and more inspectable. Keep it — but make it self-populating via the onboarding workflow.

4. Single VPS: We don’t need microservices. All of these enhancements fit within the existing Python + Docker deployment on a single VPS.

5. Workflow-driven, not autonomous chat: The PRD’s insight — “deterministic workflows prevent the agents-talking-endlessly problem” — remains correct. Even Okara structures its agents as pipelines, not freeform conversations.

---

Summary

Crystal-Agents and Okara share the same architecture. The gap is not structural — it’s in depth of marketing intelligence and UX that surfaces value proactively. The two highest-impact enhancements are:

1. Action Feed — turn the dashboard from a log viewer into a decision-making surface
2. Business Onboarding workflow — make the system learn your business autonomously instead of requiring manual memory files

Everything else follows from those two.