№ 07·0207 · AI Agent System3 min read · Section 2 of 5
7.2 Agent type
The five types of Agents map different tool permissions, log granularity and settlement calibers; they are aligned with roles in the orchestration framework, but are bound to governance.
Updated
7.2 · Agent type
Five types = five mission identities, not five chat skins.
Multi-agent frameworks such as CrewAI use “roles” to organize prompts and tools; WCN upgrades roles to governance objects: each type of Agent corresponds to different data visibility, executable actions, and PoB attribution templates. Without classification, permissions and settlement will be confused.
What this page doesType definition × Permission × Log × PoB caliber
Reading highlightsWho can see what, what tools can be adjusted, what marks can be left
Five typical Agents
Research AgentInformation retrieval, multi-source summary, structured table of comparable companies/tracks, and due diligence list filling. Tool examples: Intranet document RAG, permitted third-party data API, prohibiting direct connection to any web page crawling (anti-injection). LLM Note: Fact references must have source anchors; values are marked "to be verified" by default.
Deal AgentCapital-project match candidates, teaser-level pitch drafts, deal stage status recommendations. Prohibited: Issuing binding terms or exclusive commitments on behalf of nodes without authorization. Copilot comparison: Copilot optimizes single-person manuscripts; Deal Agent output needs to have deal_id and recommendation logic summary for review.
Growth AgentMulti-version copy, funnel bullet points, post-campaign attribution summary (tied to the Growth task). Risk: Must go through the manual release gate before publishing to the outside world; avoid combining with prompt words such as "profit commitment".
Execution AgentMeeting minutes, to-do breakdown, reminder draft, and receipt tracking list. Benchmark: Closest to enterprise automation (Zapier/Make) + LLM, but each suggestion should be linked to meeting notes and timestamps.
Liquidity AgentLiquidity and market health monitoring summary, checklist status, abnormal threshold alarm description. TradFi Lenovo: Similar to the "natural language layer" of a certain type of risk dashboard in Aladdin, the decision-making power still lies with the trading desk/node.
Why can’t they be mixed into a “super agent”?
LangChain-style general-purpose Agents often share the same toolset; in institutional scenarios, the principle of least privilege will be violated. After classification, the system can do:
Who can see whatResearch can read a wide range of materials; Liquidity can only read market prices and agreed contracts; Deal does not read sensitive fields of salary levels in the entire database by default.
Who can do whatExecution can write tasks and comments; Growth can only write marketing draft areas; no type of Agent can transfer money or sign contracts by default.
Who should keep what log?Research records the search query and fragment ID; Deal records the recommended features and versions; Liquidity records the indicator definition and data time.
How and who gets into PoBAttribution templates are split by type: for example, "Adopted Research Paragraphs" and "Adopted Growth Materials" have different Proof fields to avoid settlement disputes.
Mapping relationship with multi-agent orchestration
The agent / node in CrewAI / LangGraph can be compared to the types of WCN, but WCN fixes three additional things: (1) The tool whitelist allowed for each type; (2) The state machine with the human approval edge; (3) The output is bound to task_id, node adoption event. Without (3), no matter how beautiful the arrangement is, it is just an internal experiment.
Not classified → Permission expansion → Logs cannot be defended → PoB cannot be priced. Type is a prerequisite for governance and commercialization, not a document decoration.
Short example: same event, multiple types of collaboration
Research
Pull public financing news + internal memo to generate "Conflict Points and Question List".
Deal
Output a draft of "Next round of communication points" based on the list, with matching reasons (no automatic emailing).
Execution
Write the meeting conclusion as an action item with owner/due and attach it to Task.
Growth (optional)
Generate external neutral update copy from the adopted facts, and manually click to publish it.
The goal of types is to answer at the architectural level: if something goes wrong, is it a problem with the model, tools, configuration, or human adoption - it can be broken down before it can be managed.