№ 04·0304 · How WCN works4 min read · Section 3 of 4
4.3 The division of work between people and agents
Where people stay accountable, where agents apply, and the controls that keep agent output auditable.
Updated
4.3 · People and agents
People own direction and accountability; agents own execution and amplification.
WCN does not replace people with AI. It places human nodes and AI agents in one chain of responsibility: people own direction, resources, and final accountability; agents own execution and process automation. The boundary follows what current AI can and cannot do, not a theoretical split.
Core principleJudgment and accountability stay with the person; execution scales with the agent
Boundary basisThe real ceiling of current model capability
Control requirementEvery agent output is auditable and reversible
An agent in WCN is not a marketing feature. It is the network's execution amplifier — controlled by human judgment, logged on every action, and never the final signer.
Human nodes: four areas that stay with people
In WCN, human nodes are not targets for replacement. They are the strategic control layer. Under current conditions, four areas must stay under human responsibility.
Resource introduction and relationshipsBringing projects, capital, services, and regional resources into the network. High-value B2B relationships rest on trust, judgment, and reputation between people. An agent can assist research; it cannot replace who you know and who trusts you.
Strategic judgment and priorityDeciding what advances, what pauses, and what moves to the next stage. Current models are strong at aggregation and pattern recognition, but business judgment, market timing, and political sensitivity still need human experience.
Trust and the responsibility it carriesWhen a multi-million-dollar decision needs a signature, the counterparty weighs the recommender's reputation and legal liability, not a model's score. High-value Web3 deals remain a trust exchange between people.
Final accountability and complianceSigning contracts, sending payments, issuing legal opinions, reaching tax conclusions, and filing with regulators carry a defined legal subject. When an overstep or error occurs, responsibility falls on the node, not the model.
Industry reference. In 2024 Goldman Sachs deployed internal AI to assist IPO due diligence and contract review, yet MD-level bankers still sign every final decision. JPMorgan's contract review system processes a high volume of documents, but a lawyer confirms each legal conclusion. That is the current boundary of practice.
Agents: where they apply best
The value of an agent is not full automation. It is amplifying a human node's output within defined boundaries.
Research AgentProject summary, competitive analysis, risk marking, and market data assembly. It compresses a multi-day analyst task into minutes. It outputs a structured report that a human reviews, then adopts or revises.
Deal AgentProject-to-capital match suggestions, first-pass screening scores, a due-diligence checklist, and deal progress tracking. It does not make investment decisions; it prepares the inputs a decision needs.
Growth AgentContent drafting — project summaries, evidence-based copy, social posts — plus channel matching and attribution tracking. It standardizes the distribution capability of a BD team.
Execution AgentMeeting minutes, to-do tracking, reminders, status reports, and document-collection prompts. It removes the coordination tax that consumes 30 to 40 percent of deal time.
An agent's core role is to structure and replicate the methods of strong nodes — so the working method of one top operator can serve a hundred ordinary nodes.
The real boundary of AI capability
The division of work follows what AI can actually do, not a marketing narrative.
Where AI is strongInformation aggregation and structuring; pattern recognition and anomaly detection; translation and localization; process automation and status tracking; large-scale filtering and sorting.
Where AI is weakJudgments with legal consequence; strategic choices in multi-party games; cultural and political risk; timing that needs industry intuition; trust that rests on personal reputation.
Where AI is improving but not yet reliableLong reasoning chains, where accuracy falls past several steps; real-time fact verification, where hallucination is not fully solved; decision consistency, where the same input can yield different output.
WCN's design responseEvery agent output defaults to a recommendation, not a decision. Key actions require human approval. Agent logs are kept intact for audit and dispute review.
Control framework: what happens when an agent errs
Output reviewEvery key agent output — recommendation, report, match result — is reviewed by a human before it goes out. The reviewer can approve, revise, or reject.
Capability and task scopeEach agent carries a Capability Manifest that fixes its boundary, and a Task Contract that authorizes scope per task. Read tasks run automatically; access and transaction actions require approval.
Audit trailEvery input, reasoning step, and output is recorded. When a result is wrong, a reviewer can trace the exact step and the data the judgment rested on.
RollbackIf a problem surfaces after a recommendation is adopted, the system marks it corrected and updates every downstream decision that relied on it.
WCN's design principle for agents: a step slower beats a step wrong. In high-value Web3 deals, one bad recommendation or one leak can cost millions. An agent's value depends on staying controllable.
Boundary summary
Stays with the personSigning, payment, legal and tax conclusions, final Proof approval, high-risk compliance judgment, major resource commitment, and speaking for the network externally.
Can move to an agentResearch, screening, match suggestions, follow-up reminders, minute generation, information sorting, status tracking, first-pass attribution calculation, and monitoring.
Needs human approval firstAny external send — email or message — confirmation of key documents, transaction-structure suggestions, and actions in sensitive jurisdictions.
Must leave a complete logEvery agent read, recommendation, contact, verification, marking, reminder, and output carries a timestamp and its context.
WCN does not chase the narrative of full automation. It makes the agent a controlled, auditable, settleable execution layer — guided by human judgment, scaling the network's execution by an order of magnitude.
