№ 04·0304 · How WCN works4 min read · Section 3 of 4
4.3 Division of roles between people and agents
Based on the true boundaries of the current AI capabilities, define the irreplaceable areas of humans and the best starting points for Agents.
Updated
4.3 · Division of roles
People are responsible for direction and responsibility, and Agents are responsible for execution and amplification—two roles in the same chain of responsibility.
The key to WCN is not to “replace people with AI”, but to put human nodes and AI Agents into the same chain of responsibility - people are responsible for direction, resources and final responsibility, and Agents are responsible for execution amplification and process automation. The boundaries of division of labor are not theoretical settings, but real boundaries based on current AI capabilities.
core principlesJudgment and responsibility rest with the person, execution and amplification rest with the Agent
Boundary basisThe true upper limit of current LLM capabilities
Risk control requirementsAll Agent output must be auditable and rollable
Human Nodes: Four Irreplaceable Fields
In WCN, human nodes are not objects to be replaced by AI, but the strategic control layer of the system. The following four areas must be under human responsibility given current technological conditions:
Resource introduction and relationship buildingBring projects, capital, services, and regional resources into the network. High-value B2B relationship building still relies on trust, judgment and reputation between people. AI can assist research, but it cannot replace "who you know" and "who trusts you."
Strategic judgment and prioritiesDecide what is worth moving forward, what needs to be paused, and what should move to the next phase. Current LLMs are strong in information aggregation and pattern recognition, but still require human experience when it comes to business judgment, market timing, and political sensitivity.
Trust and Responsibility GuaranteeWhen a $5M investment decision requires signature, the other party is not looking at the AI’s recommendation score, but the recommender’s reputation and legal liability. Web3's high-value transactions are still a "human-to-human" trust game.
Final Responsibility and ComplianceSigning contracts, payments, legal opinions, tax conclusions, regulatory filings - these actions have clear legal subject requirements. Once there is an overstep of authority, error or dispute, the responsibility cannot fall on the model, but must fall on the node body.
Industry Reference: Goldman Sachs deployed internal AI tools to assist IPO due diligence and contract reviews in 2024, but all final decisions are still signed off by MD-level bankers. JPMorgan’s AI Contract Review System (COIN) processes tens of thousands of contracts, but the legal conclusion of each requires confirmation by a lawyer. This is the true boundary of the current industry.
Agent: the best point of exertion
The value of Agent is not "full automation", but to amplify the output of human nodes by 5-10 times within clear boundaries:
Research AgentProject information summary, competitive product analysis, risk marking, and market data compilation. Condensing a VC analyst’s 3 days of work into 30 minutes. Output a structured report, which can be adopted or modified after human review.
Deal AgentProject-capital matching recommendations, preliminary screening scores, due diligence checklist, and deal progress tracking. Do not make investment decisions, but prepare the information needed for decision-making.
Growth AgentContent generation (project summaries, testimonial copy, social posts), distribution channel matching, attribution tracking. Standardize a BD team’s distribution capabilities.
Execution AgentAutomatic generation of meeting minutes, to-do tracking, reminders, status reports, and material collection reminders. Eliminate the 30-40% "coordination tax" in Deal.
The core role of Agent is to streamline, structure, and replicate the "methods of excellent nodes" - allowing the working methods of a top FA to be used by 100 ordinary nodes.
The true boundaries of AI capabilities (2025-2026)
The division of labor design must be based on the real capabilities of AI, not marketing narratives:
What AI is good atInformation aggregation and structuring (90%+ accuracy); pattern recognition and anomaly detection; multi-language translation and localization; process automation and status tracking; large-scale data filtering and sorting.
What AI is not good atJudgments involving legal consequences; strategic choices in multi-party games; cultural sensitivity and political risk assessment; timing judgments that require "industry intuition"; trust transfer involving personal reputation.
AI is improving but not yet reliableComplex reasoning chain (>5-step reasoning accuracy decreases); real-time information verification (the hallucination problem is not completely solved); consistency of autonomous decision-making (same input but different output).
Design response to WCNAll Agent output defaults to "Recommendations" rather than "Decisions". Key actions require human approval. Agent logs are kept intact for auditing and dispute backtracking.
Risk control framework: What to do when the Agent goes wrong
Output reviewAll key output of the Agent (recommendations, reports, matching results) must be reviewed by humans before being sent out. Reviewers can approve, modify, or reject.
Permission classificationAgent's permissions are graded according to risk: Reading type (low risk, can be automated) → Sorting type (medium risk, regular review) → Access type (high risk, must be approved by humans) → Transaction type (automatic execution is prohibited).
audit trailEvery input, reasoning process and output of the Agent are fully recorded. When the result is wrong, you can trace back to the specific step where the problem occurred and what data the judgment was based on.
Rollback mechanismIf problems are found after the Agent recommendation is adopted, the system supports marking the recommendation as "corrected" and updating all subsequent decision-making chains that rely on the recommendation.
WCN’s design philosophy for Agent: It’s better to be a step slower than to make a mistake. In high-value Web3 transactions, a single bad recommendation or leak can cost millions of dollars. The value of an agent must be based on controllability.
Boundary summary
Must stay in personSigning, payment, legal and tax conclusion, final PoB approval, high-risk compliance judgment, significant resource commitment, speaking to external representative network.
Can be handed over to AgentResearch, screening, matching recommendations, follow-up reminders, minute generation, information sorting, status tracking, preliminary attribution calculation, monitoring and early warning.
Requires human approval before executionAny external sending (emails, messages), key material confirmation, transaction structure suggestions, operations involving sensitive jurisdictions.
A complete log must be leftAll the Agent's reading, recommendation, contact, verification, marking, urging, and output - every action has a timestamp and context.
WCN does not pursue the narrative of “AI full automation”. What it pursues is to make Agent a controlled, auditable, and settleable execution layer - under the guidance of human judgment, amplifying the network's execution capabilities by an order of magnitude.