How to Build, Configure, and Publish Custom AI Agents in the Ernie Bot Ecosystem

Creators and knowledge workers can automate repetitive tasks—such as summarizing industry research, answering specialized customer inquiries, drafting recurring reports, and qualifying inbound leads—by building custom AI agents on Baidu's official agent development platform, Wenxin AgentBuilder (agents.baidu.com), powered by the ERNIE (文心) foundation models.

Unlike standalone chatbots that only produce conversational text, ERNIE-powered custom agents combine prompt-directed personas, private domain knowledge bases (RAG), external tool plugins, and visual deterministic workflows. Once verified, these agents can be published directly to the Ernie Bot Agent Store (within the Ernie Bot / Wenxiaoyan web and mobile apps) and distributed across Baidu Search results, giving creators unprecedented organic exposure.

Here is a practical, end-to-end guide to designing, configuring, testing, and launching custom agents to streamline everyday knowledge work.

1. Choosing the Right Agent Architecture

Before writing prompts or uploading documents, identify whether your task requires conversational flexibility or rigid process execution. Wenxin AgentBuilder provides two core development modes:

Mode A: Zero-Code Conversational Agent (Prompt-Driven)

• Best for: Copywriting assistants, brainstorming partners, roleplay coaches, customer service assistants, and general domain Q&A.
• How it works: You describe the agent's objective in plain language. The platform's AI co-creator generates baseline configurations (name, description, system prompt, and opening starter queries). You then refine the prompt and attach knowledge files.
• Advantage: Takes less than 10 minutes to prototype and launch; handles conversational ambiguity gracefully.

Mode B: Low-Code Workflow Agent (Logic & DAG-Driven)

• Best for: Multi-step pipelines requiring determinism—such as competitive data aggregation, structured report generation, API data retrieval, and multi-branch decision trees.
• How it works: You design a visual workflow canvas using directed acyclic graph (DAG) nodes:
  • Input Node: Receives user prompts, parameters, or uploaded files.
  • LLM Nodes: Assigns specific sub-prompts and reasoning tasks to ERNIE model variants.
  • Selector / Branch Nodes: Directs logical flow based on keywords, sentiment, or classification rules (if/else logic).
  • Code Nodes: Executes lightweight Python or JavaScript code snippets for data cleaning, math calculations, and JSON parsing.
  • Tool / API Nodes: Calls Baidu Search or third-party webhooks.
  • Output Node: Assembles clean, formatted results (Markdown, tables, or JSON).
• Advantage: Guarantees 100% predictable execution paths, eliminating unwanted hallucination in business-critical routines.

2. Step-by-Step Configuration Guide

Step 1: Establish Persona and Core Instructions

Navigate to agents.baidu.com and click Create Agent (创建智能体). Define the core operational profile:

1. Identity & Branding: Choose a functional, recognizable name (e.g., SaaS Weekly Digest Drafter rather than a vague name like Helper), select or generate an avatar, and write a 1–2 sentence description outlining who the agent helps and what problem it solves.
2. System Prompt (Instructions): Structure your prompt using modular sections:
   • Role & Objective: Who the agent is and what standard it delivers.
   • Input Constraints: What information it must collect from the user before executing.
   • Execution Rules: Step-by-step logic, required formatting (e.g., bulleted executive summaries, markdown tables), and style constraints.
   • Negative Constraints: Explicit prohibitions (e.g., "Do not invent external statistics; if data is missing from the knowledge base, state that it is unavailable.").
3. Proactive Onboarding: Configure a welcoming opening message and 3–4 prompt starters (preset query buttons) to guide first-time users into immediate action without hesitation.

Step 2: Ground the Agent with Knowledge Bases (RAG)

For tasks involving proprietary guidelines, product manuals, research archives, or internal standard operating procedures (SOPs), pure prompt instructions are insufficient.

• Supported Formats: Upload PDF, DOCX, TXT, Excel/CSV, or provide live website URLs.
• Automated Processing: The platform parses, sanitizes, chunks, and vectorizes your source text automatically.
• Retrieval Tuning: During user interaction, the agent performs semantic vector retrieval, extracting only relevant chunks to augment ERNIE's context window. This ensures high domain precision and shields your agent from hallucinated responses. Baidu explicitly commits to isolating proprietary knowledge base files from general foundational training datasets.

Step 3: Integrate Plugins and External Tools

Equip your agent with dynamic capabilities ("eyes and hands"):

• Baidu Search Plugin: Enables real-time web retrieval so your agent can pull current news, market indices, or industry regulatory updates.
• Multi-Modal Components: Integrate text-to-image generation, automated speech synthesis (TTS) with tailored voice timbres, or interactive digital avatars for video-first interactions.
• Custom APIs & Webhooks: Connect your team’s internal endpoints—such as sending generated leads to a CRM, fetching live inventory status, or triggering an email delivery webhook.

3. Testing, Debugging, and Guardrailing

Before submitting your agent for public or team distribution, rigorous testing in the built-in Interactive Preview Sandbox is essential:

1. Single-Turn and Multi-Turn Evaluation: Test whether the agent retains context across multiple conversational exchanges, especially when the user interrupts or alters parameters mid-stream.
2. Workflow Variable Tracing: If using the workflow canvas, trigger individual test runs to inspect input/output variables at every node, verify that condition branches route correctly, and benchmark execution latency.
3. Boundary and Adversarial Testing: Deliberately enter incomplete parameters, off-topic requests, and edge-case inputs to verify that safety guards and negative constraints hold firm without leaking system instructions.

4. Publishing, Distribution, and Monetization

Once debugging is complete, click Publish (发布) to initiate deployment.

Compliance and Platform Review

Published agents undergo Baidu’s automated and human compliance reviews (typically taking 30–60 minutes). To avoid rejections:

• Ensure the agent name directly reflects its utility and avoids generic or deceptive terms.
• Avoid making unverified claims in sensitive domains (such as medical diagnosis, stock speculation, or formal legal certification) unless proper institutional credentials and compliance disclosures are attached.

Distribution Channels

• Ernie Bot Agent Store & Square (文心智能体广场): Your agent becomes discoverable within Ernie Bot (and the Wenxiaoyan consumer app) across desktop web and mobile apps.
• Baidu Search SERP Integration: One of Baidu’s most distinctive advantages. When everyday search engine users query problems related to your agent’s domain, Baidu AI Search can render your agent as an interactive rich card directly on the search results page, driving high-intent organic traffic.
• External Embedding & API Integration: Beyond Baidu's properties, creators can deploy agents via floating web widgets, embed them inside WeChat official accounts and mini-programs, or invoke them via Baidu's backend API endpoints (getAnswer) inside third-party SaaS tools.

Value Realization and Monetization

• Direct Ad Revenue Sharing: Creators can activate Baidu’s commercial ad-sharing module, which displays non-intrusive contextual recommendations alongside agent responses and shares advertising revenue with the creator.
• Lead Generation and Service Conversion: Embed custom commercial components, such as service booking forms, consultation inquiry links, or product purchase channels.
• Operational Efficiency: For knowledge workers, the primary return on investment is immediate time liberation: automating hours of manual drafting, content repackaging, and standard ticket triaging every week.

5. Practical Automation Scenarios for Knowledge Workers

Conclusion

Custom AI agents in the Ernie Bot ecosystem represent a shift from passive conversational chat to active task automation. By combining clear persona definitions, verified private knowledge bases, and structured workflow logic on Wenxin AgentBuilder, creators and knowledge workers can build specialized digital assistants that handle routine cognitive labor with precision—while unlocking nationwide distribution across Baidu’s search and conversational ecosystem.

References

• 文心智能体平台AgentBuilder | 想象即现实
• 百度文心助手 - 办公学习一站解决