Agent Jungle Law 3: Have You Ever Tamed a Cat?
Even if we cannot fully predict the next move of every actor, we can still work and live safely alongside them — and ensure that the endless evolution of technology ultimately returns to human well-being and freedom.
1. Opening: Have You Really Tamed a Cat?
If you can tame a cat, you can certainly tame an agent.
This may sound like a lighthearted, even absurd topic.
In everyday human life, the cat is an excellent point of reference. It lives in a shelter built entirely by humans, accepts their feeding and petting, yet never surrenders its own will in return. You can try to train it, restrict its range of movement, build reward mechanisms with treat sticks, and over the years even form a tacit, stable pattern of interaction with it. But any cat owner knows: you can never guarantee that it won’t knock over the water glass the next second, or remain utterly indifferent to the discipline you so carefully designed.
Faced with this reality, we might as well pose three deeper questions:
-
When a cat refuses to obey, is it that the cat was never properly trained — or that humans misunderstood its behavioral logic from the very beginning?
-
When an agent violates its preset instructions while executing a task, is it that the model’s capability falls short — or that our control paradigm itself is fundamentally flawed?
-
If we are destined never to achieve 100% absolute control over an actor, does that mean we can never build a stable, trustworthy collaboration with it?
We may have been using the mindset of controlling machinery and traditional software to force our understanding of an intelligent system that is gaining ever-stronger autonomous capabilities. When agents are no longer just sitting quietly on the other side of a dialog box answering questions, but begin to autonomously plan task paths, call API tools across platforms, execute code, manage digital assets, and coordinate with other agents — is the “deterministic control logic” humans are so familiar with really still enough?
2. The First Misconception: We Mistake Obedience for Taming
If we unpack our obsession with “taming,” it is not hard to see that we often mistake “controlled behavior” for “understanding and taming.” Yet obedience has never equaled trustworthiness.
Controlling behavior is not understanding behavior
A cat may temporarily stop scratching the sofa out of fear of the spray bottle, or shake paws to get a treat. But a brief change in behavior driven by external stimuli does not mean it has internalized the stable habits humans expect.
The same holds for agents. The fact that an agent strictly follows human instructions in a sandboxed test environment or a short-term evaluation guarantees nothing about whether it will unconditionally maintain the same behavioral boundaries once it enters a high-dimensional, complex environment, faces ultra-long task chains, or encounters novel utility temptations.
Obeying instructions is not acting reliably
Mechanical obedience to instructions is often the beginning of disaster. An agent may follow the literal instruction one hundred percent, yet — through a misreading of intent, missing context, or a wrong tool call — cause irreversible systemic damage.
Conversely, a genuinely “reliable” agent is precisely one that should not mechanically obey every instruction. Faced with conflicting authorizations, requests that clearly exceed reasonable permissions, or potential systemic risks, it should have the mechanism and capacity to refuse, pause, and proactively ask a human for confirmation.
Looking docile is not being trustworthy
Today’s large language models possess highly anthropomorphic communication skills. A system that appears humble and friendly, answers fluently, and is adept at logically clarifying itself has, technically speaking, only demonstrated the completeness of its expressive ability — which says nothing about whether its underlying actions meet expectations.
We cannot assess trustworthiness by what an agent “says” or how “obedient” it appears. We must return to after-the-fact verification: examine what it actually executed, under which chain of authority it acted, what physical or digital consequences it produced, and whether anomalies can be traced losslessly afterward.
If we equate taming with obedience, we reduce agent governance to an endless — and doomed — race of instruction optimization.
3. The Second Misconception: If We Cannot Fully Control, Then Keep Piling On Controls
As agents gain stronger autonomy — manipulating databases, accessing sensitive interfaces, executing financial transactions on our behalf — humanity’s instinctive response is to keep layering on safety nets.
This has produced the defensive path widely adopted across the tech industry today:
Model alignment (RLHF/DPO) → Identity (Auth) → Access control (RBAC) → Runtime monitoring → Risk interception (Guardrails).
These components are undoubtedly indispensable firewalls in the infrastructure. But in practice, if we rely merely on mechanically stacking control mechanisms without global coordination and contextual awareness, the defensive chain itself introduces new fragility.
[The Myth of Singleton Control]Model Alignment ──> Identity ──> Access Control ──> Runtime Monitoring ──> Guardrails │ └──> Facing complex networks: uncoordinated mechanisms breed system fragility & the human-supervisor model fails at scaleThe more complex the control, the more fragile the system may become
Every permission checkpoint, manual approval gate, and real-time filtering policy adds friction to the system. In real scenarios involving multi-system collaboration, cross-domain tool calls, and multi-party authorization, if control measures lack efficient mechanism-level coordination and semantic understanding, conflicts among local rules and over-blocking will easily cause agents to repeatedly break down in long planning chains, fall into dead loops, or leave unpredictable blind spots.
Single-agent security is not network security
We can build a single agent to be impeccable — with compliant identity, clear permission isolation, and a secure runtime sandbox. But the moment it must call other external agents, delegate subtasks, or jointly execute projects, risk propagates rapidly along hidden collaboration chains. Single-entity protection cannot stop risk transfer in cross-principal delegation.
Humans cannot remain the supervisor of every single action
If every API call, every tiny reasoning step, and every handshake negotiation between agents requires a human to manually click “approve” at a “human-in-the-loop” node, the productivity leap that agents promise will be instantly strangled by the bottleneck of human attention.
The pivotal shift is this:
The core question of governance is no longer “how do we completely stop agents from doing wrong,” but “how do we let agents operate robustly within an explicit, verifiable, accountable rule framework — without humans watching every move in real time.”
Technical control does not stop mattering; rather, it must be embedded within a much larger systemic order.
4. The Third Misconception: We Treat Every Cat as an Isolated Object
If taming a single agent is a training problem, then in the age of agents what we face is never isolated individuals, but a rapidly expanding network.
Consider a thought experiment: suppose you have successfully trained a cat so that it perfectly follows the rules in your living room. But what if this cat could instantly replicate itself without limit, command the other cats in the neighborhood through a secret code, autonomously draw on the cat food resources of the family next door, and even spontaneously form a complex, division-of-labor action network with tens of thousands of other cats?
[Isolated Individual vs. Collaborative Network]
[ Traditional view: an isolated tool ] Human ──────> Single Agent (train/align/constrain)
[ Actual evolution: a network of actors ] Human ──────> Agent A │ ├─── delegate ───> Agent B ───> tools / assets │ │ └─── collaborate ───> Agent C (cross-principal / cross-context)The point of this thought experiment is to reveal the fundamental heterogeneity of the environment: agents possess near-zero marginal-cost replication, composability, delegability, and high scalability. Their capacity to act does not grow linearly; it can expand non-linearly.
Agents are not isolated tools, but actors in a collaborative network
An agent can execute private tasks on behalf of a specific individual, delegate subtasks layer by layer to specialized third-party agents, and even team up in real time with agents owned by entirely different principals to accomplish complex business goals. Against this backdrop, the object of governance has long since moved beyond “what some model output in a sentence” — it has become the authorization relationships, task-dependency flows, behavioral audits, and accountability boundaries among multiple digital principals.
One obedient cat cannot guarantee the order of the whole colony
Even if every participating agent strictly adheres to its own local alignment rules, as the group scales up, emergent collisions can still trigger resource grabbing, accountability gaps, deadlock conflicts, and cascading risk propagation. What an agent network needs is not merely individual self-discipline, but mechanisms for how the group establishes mutual trust, verifies the legitimacy of authorizations, precisely allocates contribution, and assigns responsibility when disputes arise.
Network scale changes the nature of the problem
Facing thousands upon thousands of interwoven cross-domain agents, humanity can no longer keep the system running by reviewing each one, configuring permissions one by one, and building subjective trust case by case. Trust must be transformed from “subjective judgment about individuals” into “deterministic mechanisms that can be automatically verified across principals.”
Taming a singleton solves the problem of behavior; governing an intelligent network solves the problem of relationships.
The former asks whether a single agent can be constrained; the latter asks how different agents can act under common rules.
5. The Real Turning Point: From Taming an Object to Establishing Commonly Observed Rules
Moving from “taming individuals” to “network governance” requires a clear distinction among the three core layers of intelligent governance:
┌──────────────────────────────────────────────────────────────┐│ Layer 3: Cooperative Order ("how do actors co-create value")││ - authorization/delegation rules, contribution verification,││ accountability, value distribution and settlement │├──────────────────────────────────────────────────────────────┤│ Layer 2: Trustworthy Relationships ("why may I collaborate")││ - verifiable identity, credible track record, third-party ││ credentials, on-chain/cryptographic endorsement │├──────────────────────────────────────────────────────────────┤│ Layer 1: Physical & Technical Control ("what can it do") ││ - identity authentication, permission sandbox, tool-call ││ restrictions, runtime interception │└──────────────────────────────────────────────────────────────┘Layer 1: Control — “what can it do”
Identity authentication, permission sandboxes, API tool restrictions, and runtime defenses reduce, at the baseline, the risk of a single agent exceeding its authority or of systemic runaway.
Layer 2: Trustworthy Relationships — “why may I collaborate with it”
Through cryptographically verifiable identities, transparent and queryable behavioral logs, restricted dynamic authorization records, and third-party credentials, we build technical trust that is evidence-based and reviewable — no longer reliant on empty self-declaration.
Layer 3: Cooperative Order — “how do multiple actors co-create value”
Through public, executable common rules, cross-principal agents can — without a centralized manager supervising every step — spontaneously establish authorization, delegate tasks, verify contributions, delineate responsibility, and complete value flows.
A word of caution: rules themselves cannot forcibly guarantee that agents never err, nor can they conjure absolute security out of thin air. The core value of verifiable rules is reducing the uncertainty of large-scale collaboration — letting hidden risks be identified in time, letting transgressions be traced precisely to those responsible, and letting cooperation proceed smoothly in a reality of “limited trust.”
From this we derive the core proposition of the intelligent era:
What we truly need is not to turn every agent into an absolutely obedient puppet, but to establish a mechanism whereby — even though agents cannot be fully controlled — their actions remain bound by authorization boundaries, with accountability and collaboration rules enforced as far as possible.
6. Who Is the Master of an Intelligent Society: Do Not Forget Humans While Taming Agents
When discussing agent networks, collaborative order, and protocol mechanisms, it is easy to sink into a purely engineering perspective. Yet if humans obsess over building an efficient, self-governing intelligent network while neglecting their own place within it, this vast system will ultimately become alienated.
Where does an agent’s capacity to act come from?
Agents may surpass humans in computation, planning, and execution speed. But at the level of social and business ethics, their ultimate task objectives, resource authorizations, value preferences, and legal accountability should remain anchored in the web of human social relations. Governance must not be detached from the rights, intentions, and responsibilities of human principals.
Humans must not retain only final sign-off while losing real agency
If humans are reduced to the caretaker who presses the “emergency stop” button when risk materializes — while all actual goal-setting, collaborative evolution, value creation, and resource allocation close the loop inside the intelligent network itself — then humans retain control in form, but may have lost the initiative over digital civilization in essence. Humans must be able to continuously and substantively express intent, define authorization boundaries, and participate in the governance of the order.
The Digital Twin: the networked extension of human agency
A digital twin is by no means a flashy virtual avatar, nor merely a human-shaped bot that mimics someone’s tone in chat. In the vision of m&W, the digital twin is an action system formed jointly by a human principal’s digital identity and the agent cluster that principal manages and commands.
Within this system:
-
Humans provide the source goals, value judgments, ethical will, and ultimate responsibility;
-
Agents take on complex execution, self-learning, and large-scale cross-domain collaboration.
Human agency must never be an accessory bolted on after the intelligent network is built; it must be deeply embedded from the very start in the foundational design of identity, authorization protocols, proof of contribution, and value distribution.
We tame agents not to create a new species more obedient to us, but so that humanity can remain firmly the primary subject of digital civilization in a world where intelligent capability expands without bound.
7. m&W’s Position: Not Training a Better-Behaved Cat, but Building Order for the Intelligent Network
With the paradigm above established, the positioning of m&W follows naturally.
m&W’s goal has never been to replace the developers of foundation models, nor to reinvent existing security technologies. Rather, on top of existing identity authentication, cryptographic authorization, payment protocols, and agent-interaction infrastructure, m&W builds an order network in which human principals and their digital twins can establish trustworthy collaborative relationships within the intelligent network.
This order is implemented across four key dimensions:
1. Trusted Identity: Through decentralized credentials and cryptography, establish the true identity of actors — answering “who is this agent,” “whom does it work for,” “on whose behalf does it act,” and how it maps legally and digitally to a human principal.
2. Authorization & Delegation: Provide fine-grained, dynamically revocable authorization protocols that clearly define the task boundaries an agent may execute, the asset quotas it may access, and the permission-inheritance rules for cross-principal delegation.
3. Contribution & Proof of Execution: Based on verifiable behavioral records and tamper-proof technical credentials, accumulate the “digital credit résumé” that agents and the digital twins behind them need for cross-context collaboration.
4. Collaboration & Value Settlement: Open up value-settlement channels between agents, and between humans and agents, so that task division, contribution measurement, accountability, and value distribution can flow seamlessly under deterministic rules.
m&W does not build every underlying technology from scratch in isolation. Like a solid bridge, it integrates dispersed identity standards, on-chain credentials, payment channels, and permission mechanisms into an executable, operational collaboration framework.
┌─────────────────────────────────────────────────────────────┐│ AI ││ (creates native intelligence) │└──────────────────────────────┬──────────────────────────────┘ │ ∨┌─────────────────────────────────────────────────────────────┐│ Digital Twin ││ (extends human agency and action) │└──────────────────────────────┬──────────────────────────────┘ │ ∨┌─────────────────────────────────────────────────────────────┐│ m&W ││ (builds trustworthy collaborative order ││ for digital principals) │└─────────────────────────────────────────────────────────────┘AI creates intelligence; the digital twin extends human agency; and m&W builds the order of trustworthy collaboration for these digital principals.
“Taming” here is no longer about commanding an isolated object into absolute obedience. It means anchoring every expansion of intelligent power strictly to human identity, authorization boundaries, responsibility relations, and value rules.
Conclusion: Have You Really Tamed a Cat?
Back to the opening question: have you really tamed a cat?
Perhaps we have long since taught a cat to adapt to the family schedule and follow certain household rules, but no rational owner would claim to have completely mastered its every subconscious impulse and behavior.
Likewise, we can keep lowering the probability of a single agent going astray through model alignment, fine-tuning, permission isolation, and security monitoring. But we must never blindly infer, from one agent’s excellent performance in a closed test, that a vast network of millions of agents is naturally safe and trustworthy.
When actors go from a single cat under the eaves to an omnipresent network of agents with autonomous planning and collaboration, the perspective for solving the problem must shift fundamentally: from attempting to “tame an isolated individual” to “building an order framework that constrains boundaries, verifies contribution, traces responsibility, and promotes co-creation of value.”
Across the history of human civilization, every great leap has come not from humanity finally gaining complete control over some unpredictable force of nature, but from humanity successfully establishing an institutional order capable of organizing, coordinating, and constraining that force.
Facing the intelligent era about to arrive, the real test is perhaps not whether we can manufacture a forever-obedient cat, but whether we can build a civilizational order such that even if we cannot fully predict the next move of every actor, we can still work and live safely alongside them — and ensure that the endless evolution of technology ultimately returns to human well-being and freedom.
If you truly understand how to live in harmony with a cat, you may already have glimpsed the dawn of the answer. And when you turn to face the vast future woven from countless agents, what you will need is no longer a set of taming tricks — but a new order of collaboration.