# Building a Self-Improving Multi-Agent System on Purple Fabric

# A Real-World Walkthrough: Automating Company Intelligence Reports

- Tarun Shah | Data Engineer

Agentic AI is moving from one-shot text generation to self-evolving systems — AI that acts, observes, and improves with every loop.
This post walks through how we built a self-improving marketing-intelligence agent system on Purple Fabric, leveraging its no-code orchestration, tool integration, and multi-agent capabilities.

The use case:
Given a company name and website, automatically discover public data (like annual reports and leadership pages), extract structured insights, assess completeness, and iterate until a high-confidence marketing report is ready for delivery.

---

# Why We Built It This Way

Generating a one-off company summary with an LLM is easy.
But for enterprise use, accuracy, coverage, and traceability are non-negotiable.
We needed:

• Verified facts (CEO, HQ, revenue, ESG metrics)
• Evidence with source URLs and document provenance
• Self-review and correction without human babysitting

So we designed a multi-agent architecture where each agent has a narrow, auditable responsibility — all orchestrated inside Purple Fabric’s visual builder.

---

#

#

# Why “self-improvement” matters for agents

Traditional single-shot LLM calls often produce useful but incomplete/partial results. Agentic systems (in Purple Fabric) solve more complex tasks by chaining “digital experts” and tools. Self-improvement adds an iterative review and directed follow-up mechanism so your system can:

• catch missing facts or low-confidence outputs
• refine strategies (eg: choose a different parsing tool)
• In research, iterative self-feedback and small adjustments have been shown to meaningfully raise task-success and robustness.
  The “think” tool is better suited for use cases which require complex tools, analyze tool outputs carefully in long chains of tool calls, navigate policy-heavy environments with detailed guidelines, or make sequential decisions where each step builds on previous ones and mistakes are costly.

---

# System Architecture in Purple Fabric

At a high level, the workflow runs as a loop of agents coordinated by a main Orchestrator Agent:

Web Discovery → Document Harvester → Evidence Extractor → Think & Audit
↳ (if incomplete) → Remediation Loop → Ready → Report Synthesizer → Compliance & Finalizer

Each agent is a Digital Expert inside Purple Fabric, connected through the orchestration canvas.

---

#

# 🔹 Agent A: Orchestrator

The Orchestrator is the control center.
It:

• Receives {company_name, website_url} as input
• Invokes all agents in sequence
• Evaluates Think & Audit’s feedback
• Triggers remediation cycles (up to two) if coverage is missing
• Packages final results for reporting or further review

In Purple Fabric, this is defined as the primary workflow agent, using condition-based routing in the canvas.

---

# 🔹 Agent B: Web Discovery

This agent finds all the starting material:

• Runs search_web (a custom Python tool integrated into Purple Fabric) using queries like
  "{company_name} investor relations site-{domain}"
  (note the enforced “site-” syntax to prevent malformed search strings)
• Crawls the main domain with crawl_site to collect additional IR, leadership, and ESG links
• Returns a ranked list of URLs — PDFs, HTML pages, and news sources

Output:
{
"company_name": "TechNova Inc.",
"ranked_urls": [
"https://technova.com/investor/Annual-Report-2024.pdf",
"https://technova.com/about/leadership",
"https://technova.com/sustainability"
],
"credible_count": 3
}

These URLs flow directly into the Document Harvester agent.

---

# 🔹 Agent C: Document Harvester

Document Harvester turns raw URLs into structured, machine-readable text.

It:

• Checks file types:
  • PDFs → processed by pdf_to_markdown and extract_tables
  • HTML → processed by html_to_markdown
• Assigns roles (annual_report, leadership_page, esg_page, etc.)
• Produces Markdown-formatted text + extracted tables with metadata

Example output snippet:
{
"role": "leadership_page",
"source_url": "https://technova.com/about/leadership",
"markdown": "# Leadership Team\nJane Smith – CEO\nMark Davis – CFO\n...",
"meta": {
"content_type": "text/html",
"status": 200,
"extraction_time": "2025-10-26T12:34:56Z"
}
}

All documents are accumulated across iterations for downstream extraction.

---

# 🔹 Agent D: Evidence Extractor

The Evidence Extractor parses the harvested markdown and emits factual statements as “evidence atoms”:

• Fact (text)
• Fact type (financials, leadership, ESG, positioning)
• Confidence score
• Source URL and anchor

Example:
{
"fact": "Jane Smith is the Chief Executive Officer.",
"fact_type": "leadership",
"source": { "url": "https://technova.com/about/leadership" },
"confidence": 0.95
}

The extractor also supports "sanitize_output": true — added to handle JSON escaping issues (Invalid \escape) and ensure robust outputs even with malformed URLs or special characters.

---

# 🔹 Agent E: Think & Audit

This is the self-improvement engine — the agent that “thinks.”

It:

1. Evaluates whether all required sections are covered (Company Profile, Financials, Leadership, Market Positioning, ESG)
2. Scores:
   • coverage_score
   • confidence_score
   • freshness_ok
3. Identifies missing or low-confidence sections

Generates a remediation_plan like:
{
"needs_remediation": true,
"recommended_actions": [
{
"target_agent": "Document Harvester",
"reason": "Leadership information missing",
"needed_sections": ["leadership"],
"hint_queries": ["{company_name} leadership team site-"]
}
]
}

4. The Orchestrator reads this plan and re-invokes the right agents automatically

---

# Remediation Loop

This is where self-improvement comes to life.

• The Orchestrator runs up to two remediation cycles
• In each:
  • Unharvested URLs relevant to missing sections are prioritized
  • Evidence Extractor re-runs with a focused context and sanitize_output:true
  • Think & Audit evaluates again

If coverage becomes acceptable → move forward.
If not → stop after 2 loops and mark status = "NEEDS_REMEDIATION".

This closed feedback loop — implemented directly within Purple Fabric’s orchestration canvas — allows autonomous refinement without external code.

---

# 🔹 Agent F: Report Synthesizer

Once Think & Audit signals readiness, the Report Synthesizer agent composes the structured marketing report:

• Executive Summary
• Company Overview
• Financial Highlights
• Market Positioning
• Leadership & Organization
• ESG & Risk
• Notes on Data Quality

All statements include citations from evidence atoms.

---

#

# 🔹 Agent G: Compliance & Finalizer

The final agent ensures enterprise readiness:

• Validates language and compliance (e.g., “the company expects” instead of “will”)
• Attaches disclaimers
• Flags confidential or unverifiable statements
• Packages final JSON output with compliance status and metadata

Output structure:
{
"status": "READY_FOR_DELIVERY",
"final_report": ,
"citation_index": [...],
"compliance_status": {
"approved_for_distribution": true,
"risk_notes": []
},
"packaging_metadata": {
"coverage_score": 0.89,
"confidence_score": 0.92
}
}

---

# How Self-Improvement Works Inside Purple Fabric

The beauty of this implementation is that everything happens inside Purple Fabric, no external orchestration code required.

This pattern forms a genuine agentic feedback loop:

Act → Observe → Think → Re-Act → Validate → Deliver

---

# Results and Observations

• Autonomy: The system can run end-to-end for most companies without manual intervention
• Precision: By the second loop, coverage scores improved from ~0.6 → ~0.9 on average
• Resilience: The Think & Audit + remediation design gracefully recovers from missing data or parser errors
• Traceability: Every fact is tied back to a source document and extraction timestamp

---

# Why Purple Fabric Was Ideal

Purple Fabric enabled us to do all this without writing custom orchestrator code:

• Each agent (Digital Expert) was designed in the no-code canvas
• Each Python tool (search_web, crawl_site, pdf_to_markdown, html_to_markdown) was integrated as reusable “Custom Tools”
• The feedback loop logic (audit → remediation → retry) was embedded directly in the Orchestrator’s system prompt — not hardcoded externally
• Governance, observability, and approvals were included automatically

Essentially, we implemented a self-healing, knowledge-driven multi-agent system that stays compliant, explainable, and continuously improving — all within Purple Fabric.

---

# Summary

Agentic AI isn’t just about chaining tools.
It’s about closing the loop — teaching agents to reflect, plan, and act until confidence is high.

Using Purple Fabric, we built a fully orchestrated pipeline where:

A company name and website → trigger
Web discovery → document harvesting → evidence extraction → self-audit → remediation → synthesis → compliance → final delivery

Every phase improves the previous one.

This is self-improvement by design — agentic intelligence that learns, refines, and delivers enterprise-grade outputs reliably and transparently.