Multi-Agent Content Generation — Standard vs Deep Research Mode

What Is Multi-Agent Content Generation?

Most AI content tools call a single language model once and hand you the output. Citadex doesn't. Behind every article we generate, a pipeline of six specialised AI agents takes turns — researching the topic, planning the structure, writing, editing, optimising for AEO, and headlining — each handing off to the next.

The advantage is concrete: a single-shot LLM call is fast but tends to produce generic content with fabricated statistics and weak SEO structure. A multi-agent pipeline produces articles that are researched, structured, sourced, and AEO-ready — which is what AI engines actually cite.

You choose between two pipeline modes when generating: Standard (faster, lighter) or Deep Research (slower, more thorough). This article covers what each does, when to pick which, and what to expect from the output.

The Two Modes at a Glance

How to Pick a Mode

Open Citadex → Prompt Tracking → click Generate Strategy Article (the banner above the prompt table).

In the modal that opens, look for the "AI Agents" toggle at the bottom — turn it ON, then choose between Standard and Deep Research.

A quick rule of thumb:

• You're writing for a single prompt cluster (3–5 related queries) → Standard is usually enough
• You're writing a comparison article (Citadex vs X), a "best of" listicle, or a flagship guide → use Deep Research
• You're producing 5+ articles in a week to fill content gaps → Standard for most, Deep Research for the 1–2 hero pieces
• The topic is in your Knowledge Base already → Standard works well (RAG fills the research gap)
• The topic is broad/new and your KB is sparse → Deep Research, because Researcher needs the web

What Each Agent Actually Does

Researcher

Pulls together everything the Writer needs to know:

• Both modes: Retrieves up to 6 relevant chunks from your Brand Knowledge Base. Reads Claude's training knowledge on the topic.
• Deep Research only: Performs up to 3 live web searches using Anthropic's web tool. Brings back current sources, statistics, and competitor positioning.

Outputs a structured research brief — claims, sources, examples — that downstream agents use as ground truth.

Strategist (Deep Research only)

Plans the article. Decides:

• The target reader and what they actually want from the topic
• The article shape (listicle / comparison / how-to / standard)
• Section structure and angle for each section
• AEO hooks — what makes this piece citable by AI engines
• Which key data points from the research brief deserve emphasis
• Also retrieves up to 4 competitor chunks from your Knowledge Base for positioning awareness

Without the Strategist, the Writer just dumps research into prose. With it, the article has a deliberate arc.

Writer

Drafts the article body. Always grounded in:

• The Researcher's brief
• The Knowledge Base chunks (your brand facts, your real pricing, your case studies)
• The Strategist's plan (Deep Research mode)

The Writer follows hard rules baked in by us: no fabricated statistics, no soft-sell CTAs in the body, brand mentions capped at 2 per article (unless it's an explicit comparison piece), proper citation of every data point.

Editor (Deep Research only)

Reviews the Writer's draft against a publish-safety checklist:

• Removes or rewrites any statistic without a clear source
• Trims excess brand mentions
• Strips inline CTAs ("contact us today" etc.)
• Tightens prose, fixes flow

If you've ever wondered why Deep Research articles read more polished — this agent is why.

SEO Optimizer (Deep Research only)

Adds the layer that makes content AEO-friendly:

• Internal links to relevant pages on your site (using the KB to pick real targets)
• Schema markup suggestions (Article, FAQPage, BreadcrumbList)
• Meta description optimised for click-through
• Refined H2/H3 hierarchy

Headline

Generates 3–5 headline candidates and picks the strongest. Considers click-through, keyword match, and emotional pull. You can still override in the editor after.

What You See After Generation

Both modes drop the result into Content Drafts as a draft article. From there:

1. Preview — read the full article in-app
2. Edit — open in the inline editor to tweak anything (markdown supported)
3. See sources — expand the article details to see which KB chunks grounded it and which web sources the Researcher cited
4. Approve — moves the article to "ready to publish"
5. One-click distribute — publish to your blog (via GitHub integration) and syndicate to Medium / Substack / dev.to / Hashnode / LinkedIn etc. via the browser extension

Cost and Quota

Every project has a monthly quota of generations. Modes consume different amounts:

• Standard: 1 quota credit per article
• Deep Research: 5 quota credits per article (reflects the cost of web search + 6 agents)

You can see your remaining quota in the Strategy Article Modal under the mode toggle. When you hit zero, generation is blocked until the next billing cycle — or you can upgrade your plan from the pricing page.

The Quick mode also runs on cheaper models internally (Haiku for Researcher), so on a per-credit basis it's even better value than the headline ratio suggests.

When Generation Fails Mid-Pipeline

Multi-agent pipelines can fail at any stage — a web search may time out, the Writer may exceed its token budget, etc. When that happens:

• Standard mode: retries the failed agent up to 2 times silently. If it still fails, you'll see an error in Content Drafts with the stage that broke.
• Deep Research mode: because each agent's output is persisted as it completes, a mid-pipeline failure preserves prior work. You'll see a partial draft with a note like "Editor stage failed — Writer output available."

You won't be charged a quota credit for failed generations.

Best Practices

1. Fill the Knowledge Base first. Both modes use RAG. Articles generated against an empty KB read generic; articles generated against a well-stocked KB sound like your team wrote them.

1. Don't generate the same prompt twice in a row. The pipeline doesn't aggressively de-dupe topical overlap. Two articles on the same prompt cluster will compete with each other on your blog. Use Prompt Tracking to see what you've already covered.

1. Use Deep Research for comparisons and pillar pieces. Anything where the article will rank for "best X" or "X vs Y" should be Deep Research — the Strategist's competitive awareness is the entire point.

1. Review before publishing, always. Even the best multi-agent output benefits from a human read-through. Catch personality drift, factual edge cases, and brand-voice alignment.

1. Keep an eye on the agent trace. Both modes log which agents ran and what they returned. If an article reads odd, the trace usually shows where (often it's the Researcher's brief being thin because a web search returned little).

Frequently Asked Questions

Q: Can I run just one agent (e.g., just the Headline agent on existing copy)?

Not yet through the UI. The full pipeline is what's exposed; individual agents are not currently runnable in isolation.

Q: Does Standard mode skip the Knowledge Base?

No. Both modes use RAG. The difference is that Standard mode skips the live web search (Researcher only uses Claude's training + your KB) and skips the Strategist / Editor / SEO Optimizer steps.

Q: Why does Deep Research take so long?

Each web search is a real HTTP round-trip, and each agent step waits for the previous one to finish. A typical Deep Research run is: 30s research + 30s strategy + 90s writing + 30s editing + 20s SEO + 10s headlines = 3–4 minutes minimum.

Q: Can I cancel mid-generation?

Yes, from Content Drafts. Cancelling refunds the quota credit. Already-completed agent outputs are saved so you can salvage a partial draft if needed.

Q: How do I make Deep Research output longer / shorter?

Length is mostly determined by the prompt cluster shape and the Strategist's plan. To bias toward longer output, select more prompts (3–5) when launching Strategy Article. To bias shorter, select just 1–2.

Q: My output reads too "AI-generated." What now?

Two fixes, in order of impact: (1) Add more sources to your Knowledge Base — generic prose comes from a thin KB. (2) Use Deep Research mode — the Editor stage specifically rewrites bland passages. If neither helps, the prompt cluster itself may be too vague — pick prompts that are more specific.