LOOP™ (Living Oversight & Operations Protocol) is the AI agent governance protocol designed by Koneetiv. It assigns every action an agent can take to one of 4 trust zones, defines 3 escalation levels and keeps a living registry of agents, so that the autonomy granted to the machine matches the actual risk of each action.

Why a new governance protocol?

Agents act where chatbots only answered

Agentic AI changes the nature of risk. With a chatbot, the worst drift is a bad answer. With an agent, the worst drift becomes an executed action: a wire transfer sent, a contract signed, an HR decision made. Governance frameworks designed for passive tools were never built to handle that shift.

Oversight requirements are now written down

Most organizations are starting from behind. According to Deloitte (State of AI in the Enterprise, 2026 edition), only 21% of companies have mature governance of their agents. Meanwhile, ISO 42001, the EU AI Act and the NIST AI RMF spell out concrete expectations on traceability, oversight, documentation and reversibility. LOOP™ was designed to meet them without slowing execution down.

The stakes go beyond compliance. In June 2025, Gartner predicted that over 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. A governance protocol addresses the lack of risk controls directly, and it also makes teams state the expected business value from the start, since every agent needs a business sponsor accountable for it.

LOOP™ exists so that teams can delegate with confidence: everyone knows in advance what the agent does alone and what goes to a human.

The 4 trust zones

LOOP™ assigns each action an agent can take to a zone, based on the risk level, the impact of an error and the reversibility of the action. Each zone also maps to an agent confidence threshold: above 90% for green, 70 to 90% for orange, below 70% for red.

Green zone: autonomous execution

The agent acts alone, without prior approval. Its actions are logged continuously and checked afterwards by sampling. Typical uses: document sorting, email categorization, ticket routing. This zone is reserved for low-risk actions that a human can undo without harm if a sample reveals an error.

Orange zone: human approval before execution

The agent prepares the action and submits it to a designated person, who approves it before it runs, within a response window set for each use case. Typical uses: a customer reply, candidate pre-screening, a legal summary. Because nothing goes out without approval, an error is caught before it reaches anyone outside the organization.

Red zone: mandatory escalation

The agent stops. It documents what it understood and what it is missing, then alerts the owner named in the registry, who must decide within 4 hours: no action runs automatically in the red zone. Typical cases: legal decisions with contractual implications, financial commitments outside the agreed scope, unforeseen ambiguous situations, unexpected sensitive data. The agent still does useful work upstream by preparing the analysis the human decides on.

Black zone: immediate block

The black zone covers everything outside the agent's scope. As soon as the agent detects such a situation (access to data beyond its classification, an attempt to bypass its guardrails, a prompt injection), it halts execution, alerts the CISO and triggers a full audit trail. Organizations also list here the decisions they refuse to hand over to an agent, such as a medical diagnosis, a dismissal decision or a disciplinary sanction, and the governance committee reviews that list every quarter.

To classify your own use cases, our guide to classifying AI agents across the trust zones goes through the criteria one by one.

The 3 escalation levels

LOOP™ expects the agent to trigger each escalation level on its own:

In practice, the hard part is setting the thresholds. Set them too low and approvers drown in requests until they approve without reading; set them too high and actions go out that nobody has looked at. Thresholds should be calibrated on observed data and adjusted at every review.

The living registry

LOOP™ requires a single registry of every agent deployed in the organization. For each agent, it records identity and version, the named owners (business, technical, escalation, security), the data the agent can access and how that data is classified, its instructions versioned like code, its scope and blocking conditions, its operating metrics, and the history of its incidents and updates. The registry is open to the CAIO, the DPO, the CISO and internal audit, and it settles any dispute about what an agent may do.

Regulatory alignment

LOOP™ is aligned with ISO 42001 (AI management system), the EU AI Act (risk-based classification) and the NIST AI RMF (traceability and monitoring). The registry, the action logs and the recorded escalation decisions give auditors a body of evidence that already exists when they ask, and nobody has to reconstruct after the fact who approved what.

Practical implementation

Rolling out LOOP™ follows a stable sequence, whose pace depends on the number of agents, data maturity and team availability:

  1. Inventory and classification: audit existing use cases, then assign each action to an initial zone.
  2. Ownership and rules: name the sponsors, set the approval windows, write the escalation rules.
  3. Registry and monitoring: set up the registry and the monitoring tooling.
  4. Training and dry run: train the teams, then hold a first dry-run quarterly review.
  5. Adjustment and go-live: correct the thresholds in light of the dry run, then move to production.

A common mistake is to classify by agent rather than by action, which drags the whole agent into the zone of its riskiest action and strips out most of its value. Another is to write escalation rules in a document the code never reads, whereas an escalation rule has to be enforced by the agent itself to have any effect.

How Koneetiv supports the rollout

Koneetiv supports implementation through Claude Cockpit, its outsourced Chief AI Officer service with LOOP™ governance built in. On the IT side, CIO Boost also sets up an AI code governance policy submitted to the CISO. The protocol's description is public, but applying it takes a working knowledge of the regulatory frameworks and of each business function's constraints.

LOOP™ in practice: a customer collections agent

To make the protocol concrete, take an illustrative scenario: an agent that chases overdue invoices for a finance department. The thresholds and durations below are an example configuration, which each organization adjusts to its own situation.

Step 1: Initial classification

The agent identifies overdue invoices, drafts a personalized reminder and sends it. First analysis: a reminder email commits the company to its customer. Reversibility is medium (you can apologize) and so is the impact, since the customer relationship is at stake. Starting classification: sending the reminder belongs in the orange zone.

Step 2: Defining escalation rules

The configuration then splits the action by amount, which amounts to applying several zones to the same agent. Below €5,000, the email goes out automatically. Between €5,000 and €50,000, it waits four hours for the accountant's approval; past that window, the reminder stays on hold and the owner is alerted. Above €50,000, the CFO must approve. If the customer has less than six months' history or a commercial dispute is open, sending is blocked.

Step 3: Progressive deployment

During an observation phase, set here at two weeks, the agent runs in shadow mode: it drafts the emails without sending them, and the accountants compare them with what they would have written. If the comparison is conclusive at the end of that phase, the agent moves to live production on a limited scope.

Step 4: Monitoring and evolution

After a period of stable, incident-free operation whose length was set at the outset, the review committee re-evaluates the agent. If it defined an exit criterion up front, for instance 98% of cases handled without human correction, and the agent meets it, the committee can move some categories to the green zone, such as reminders between €5,000 and €20,000 sent to customers with more than two years' history, which would then go out without waiting for the accountant.

LOOP™ and the organization

A governance protocol only holds if named people carry it. LOOP™ splits responsibility for each agent between the business, AI leadership and IT.

The business sponsor

The sponsor owns the value expected from the agent and its operational responsibility. They approve thresholds and exceptions and arbitrate when an incident occurs, because it is their operation that bears the consequences.

The CAIO (Chief AI Officer)

The CAIO keeps the agent portfolio coherent, aligned with the regulatory framework and properly recorded in the registry. The role can be held internally or outsourced through Claude Cockpit, a service designed for mid-market companies without dedicated AI staff.

The IT lead

The IT lead owns technical integration, security and production monitoring, and controls the shutdown mechanism, which they trigger together with the CISO when an incident calls for it.

The governance committee

None of these roles changes an action's zone alone. Any change goes through a committee that brings them together, and every decision is justified and then recorded in the registry, which guards against reclassifications of convenience.

Pitfalls that LOOP™ helps avoid

The protocol was built to prevent mistakes that come up from one agent rollout to the next:

  1. An agent deployed without a clear business sponsor
  2. Classification that shifts with the committee's mood
  3. Escalation rules missing from the code
  4. An agent registry that nobody keeps up to date
  5. Monitoring limited to technical metrics, with no business indicators
  6. No quarterly review
  7. Untracked agent updates
  8. Dependence on a single model provider
  9. Misalignment between IT, the business and compliance
  10. No shutdown strategy in case of incident

An agent without a documented shutdown procedure should not go live. Article 14 of the EU AI Act (Regulation (EU) 2024/1689), on human oversight, requires that the people overseeing a high-risk system can interrupt it through a stop button or a similar procedure.

LOOP™ versus other market frameworks

Several frameworks coexist in the AI governance market, each with its own logic and scope. LOOP™ sits at the operational level, where each agent's behavior is actually set.

ISO 42001: the management system

ISO 42001 defines an AI Management System. It is a certifiable standard focused on organization, roles and processes, and it leaves each company free to choose its operational mechanisms. LOOP™ fits within it as an execution protocol and produces part of the evidence an auditor asks for, starting with the inventory of AI systems.

EU AI Act: the regulation

The EU AI Act classifies AI systems by risk level (unacceptable, high, limited, minimal). LOOP™ carries that classification logic into its 4 zones and applies it action by action. The mapping is partial: the black zone covers at a minimum the practices prohibited under Article 5 of the regulation (unacceptable risk), and an organization can add decisions the text classifies as high-risk, such as dismissal decisions (Annex III), if it refuses to hand them to an agent.

For high-risk systems, two obligations in the regulation overlap directly with the protocol: the automatic recording of events (logs) under Article 12, which the LOOP™ registry and action logs feed, and human oversight under Article 14, which the orange and red zones organize. The AI Act application timeline determines when those obligations apply to each system.

NIST AI RMF: the risk management framework

The NIST AI Risk Management Framework, a voluntary US reference, structures risk management into four functions: Govern, Map, Measure, Manage. LOOP™ covers all four through its own four loops: Observe, Orient, Optimize and Prove.

Consulting firm frameworks

Several consulting firms have formalized their own AI governance framework. LOOP™ is defined by its operational level of detail: each action assigned to a zone, confidence thresholds, escalation levels and a living registry. These mechanisms are described on the page dedicated to the LOOP™ protocol, so an IT department or a risk committee can examine its logic before any commitment.

How to start with LOOP™

The right starting point depends on the organization's maturity, and several paths lead to the protocol:

  1. Read and self-apply: study the public description of the protocol, then classify your own agents.
  2. Claude Cockpit guidance: a rollout guided by a Koneetiv expert, paced by your agent portfolio.
  3. Full integration via Claude Ops: tooling, protocol and monitoring, within an end-to-end agent deployment program.

Before choosing, the AI maturity assessment helps locate your starting point: it scores governance as one of its six axes and places your organization in a few minutes.

What LOOP™ changes for executives

For a leadership team, LOOP™ first provides a shared vocabulary for IT, the business and compliance, who stop debating "AI" in general and start discussing classified actions. The living registry then gives a continuous view of the agent portfolio. Because the framework is set in advance, decisions on new use cases move faster, and the company can show regulators, auditors and customers how its agents are controlled.

That last point can also help in procurement: when a tender asks a supplier about its AI governance, a structured protocol provides a documented answer, drawn from the registry and the action logs.

The Governance page sets out the full protocol, and our analyses of AI governance and compliance go further into the topics covered here.