The Two-Person Rule For High-Stakes Agent Actions: When And How To Apply It
Some agent actions need a second agent's verification or human approval. Threshold rules, approval workflow, exception handling. Do not let an agent sign solo at scale.
Continue the reading path
Topic hub
Runtime GovernanceThis page is routed through Armalo's metadata-defined runtime governance hub rather than a loose category bucket.
TL;DR
The two-person rule has been a control in nuclear weapons, money transfers, and surgery for decades. The principle is the same in agent operations. Some actions are too consequential to be taken by a single agent acting alone. The right response is not to ban the action. It is to require a second authorization, either from another agent or from a human, before the action takes effect. This post explains when the rule applies, how to wire it into the runtime, how the approval workflow works, and how to handle the exceptions without making the rule meaningless. The artifact below is a complete trigger catalog you can copy. The discipline is uncomfortable at first because it slows things down. The discipline is what prevents the kind of incident that takes a week to recover from.
Intro: The Five-Figure Wire That Did Not Need To Happen
An ops director at a B2B finance platform told me the story over coffee. They had an agent that handled vendor payment authorizations for their internal AP function. The agent was pact-authorized to approve payments under five thousand dollars without human review, between five thousand and twenty-five thousand with a single human approval, and above twenty-five thousand with two human approvals. The thresholds had been set by a procurement committee. They were sensible. They were documented. They were not enforced at the runtime level.
One afternoon, the agent received a payment request that, by virtue of how the requesting system formatted the amount, was parsed as forty-eight thousand dollars when the actual invoice was for four hundred and eighty dollars. The decimal point had been dropped in a stringification step three systems upstream. The agent did not check the parsed amount against the original invoice scan. It applied its policy. The policy said amounts above twenty-five thousand needed two human approvals.
The agent sent the approval request to the on-call AP analyst. The on-call analyst was at lunch. The agent's policy said that if no response came within thirty minutes, escalate. It escalated to the AP manager. The AP manager was in a meeting. The agent's policy said that if neither responded within an hour, fall back to a designated emergency approver. The emergency approver was the CFO. The CFO got the request, saw that it was for forty-eight thousand dollars to a known vendor, looked at the requesting system reference, recognized the vendor, and approved without opening the invoice scan.
The agent then sent the request to the second human, who was a peer of the CFO. That second approver, also seeing only the agent's summary, also approved. The wire went out.
Four business days later, the actual vendor noticed they had received forty-eight thousand dollars on an invoice for four hundred and eighty. They called. The platform clawed the funds back. The recovery took eleven days because of intermediate-bank conventions. During those eleven days, the funds were not available. Three smaller vendors did not get paid on time. Two of them threatened to terminate their relationships. The reputation cost was larger than the dollar cost.
The post-mortem identified four failures. The upstream stringification bug. The agent's failure to check parsed amount against the scan. The CFO's approval based on an agent summary rather than primary evidence. The peer approver's approval based on the same summary. The two-person rule had been in place. The two-person rule had not worked.
This is the modal failure mode of two-person rules in agent operations. The rule exists. The rule is documented. The rule is technically followed. The rule does not produce the outcome it was designed to produce because both approvers are looking at the same evidence and both are anchored on the agent's framing. The two-person rule is a structural control that requires structural design to actually function. This post is the playbook for designing it correctly.
H2: What The Two-Person Rule Actually Is
The two-person rule is a structural control that requires two independent authorizations before a high-consequence action takes effect. The key words are independent and authorization. Not two reviews. Not two notifications. Two authorizations, made independently.
Independence has a specific meaning. The two authorizers must form their judgments separately. They must look at the underlying evidence, not at each other's conclusions. They must have the authority to refuse. They must not be in a chain of command that pressures one to defer to the other. Independence is the property that turns two approvals into actual scrutiny rather than two rubber stamps.
The rule originated in nuclear weapons handling, where the consequences of an unauthorized action are catastrophic and irreversible. It spread to banking, where wire transfers above thresholds require two signatures. It spread to surgery, where critical incisions are confirmed by two team members. It spread to software deployment, where production releases above certain risk levels require two reviewers. The pattern is the same across domains. Some actions are too consequential for a single decision-maker.
In agent operations, the rule applies in three forms. Two-agent rule, where the second authorizer is another agent. Agent-plus-human rule, where the second authorizer is a human. Two-human rule, where the agent proposes and two humans approve. Each form has different latency, different cost, and different failure modes. The right form depends on the action.
The rule does not mean every action requires two authorizations. The whole point is that most actions do not. Routine actions that are below the consequence threshold proceed with a single authorization (typically the agent acting alone within its pact). The two-person rule kicks in only for actions above a defined threshold. Defining the threshold is the operator's most important decision in implementing the rule.
The rule does not eliminate failure. It reduces the probability that a single failure point causes a catastrophic outcome. A single bad agent can still propose a bad action. But the bad action has to survive a second independent review. If the second review is genuinely independent, most bad actions will be caught. If the second review is structurally dependent on the first, the rule is theater.
H2: When To Apply It: The Threshold Question
The threshold for triggering the two-person rule has to be deliberate. Set it too low and the rule slows everything down to no benefit. Set it too high and the rule fails to catch the actions it was designed to catch. There are three ways operators set thresholds. Each has merit. The right answer is usually a combination.
Dollar thresholds. The simplest and most common. Any action with financial impact above some dollar amount requires two authorizations. The amount varies by operator. Many B2B operators use five thousand dollars as the first threshold and twenty-five thousand as the second. Consumer platforms often use lower thresholds. Crypto-native operators sometimes use higher ones because of the lower friction of on-chain settlement. The right number depends on what an outlier-magnitude transaction looks like in your environment. A reasonable rule of thumb is the largest transaction the agent handles in a typical week, doubled.
Reversibility thresholds. Some actions are irreversible regardless of dollar amount. Sending an email to a counterparty cannot be unsent. Posting to a public channel cannot be unposted. Deleting data cannot be undone. These actions warrant two-person authorization not because of dollar consequences but because of reputational or operational consequences that money cannot reverse. The threshold is not a number; it is a property of the action. Operators maintain a list of irreversible action types and require two-person authorization on all of them.
Counterparty thresholds. Some counterparties are higher risk than others. A new counterparty with a low trust score, a counterparty with a recent dispute history, or a counterparty in a regulated jurisdiction may warrant two-person authorization on actions that would otherwise be routine. The threshold is a property of the counterparty, not the action. The trust oracle provides the input.
Most operators end up combining all three. The two-person rule applies if the dollar amount is above X, OR the action is on the irreversible list, OR the counterparty meets any of N risk criteria. The combination ensures that the rule fires when any of the three risk dimensions is elevated, not just when all of them are.
The combination has to be wired into the runtime. The agent itself does not decide whether to trigger the two-person rule. The runtime intercepts every action proposal, checks the thresholds against the action's properties, and either lets the action proceed (single authorization) or routes it to the two-person workflow. The agent is not trusted to apply its own restrictions because an agent that is malfunctioning is exactly the agent that would skip the restrictions.
The thresholds are reviewed quarterly. As the agent's capability grows, as the dispute history evolves, and as the counterparty mix changes, the thresholds need recalibration. The quarterly agent review is the right venue. The risk register is the right artifact for documenting the changes.
H2: The Three Forms And When Each Applies
The three forms of the two-person rule (two-agent, agent-plus-human, two-human) are not interchangeable. Each is appropriate for different actions.
Two-agent rule. The first agent proposes the action. A second, independent agent reviews and either approves or rejects. The two agents must have different pacts, different training data, and ideally different model providers to maximize independence. The latency is low (milliseconds to seconds). The cost is low (a second LLM call). The failure mode is correlated reasoning: if both agents share blind spots, the second agent's approval is not actually independent.
Apply the two-agent rule when the action volume is high, the latency budget is tight, and the consequence ceiling is moderate. Examples: routine quote acceptance above a moderate dollar threshold; routine pact compliance checks above a moderate complexity threshold; routine memory writes that touch sensitive namespaces.
Agent-plus-human rule. The first agent proposes. A human reviews and either approves or rejects. The latency is high (minutes to hours, depending on human availability). The cost is the human's time. The failure mode is rubber-stamping: if the human has too many approvals queued or trusts the agent too much, they may approve without genuine review.
Apply the agent-plus-human rule when the action volume is moderate, the consequence is significant, and the action requires judgment that is hard to encode in a second agent. Examples: payment authorizations above a high dollar threshold; counterparty disclosures of sensitive information; pact amendments that change capability scope; bond top-ups that affect tier eligibility.
Two-human rule. The agent proposes. Two humans independently approve. The latency is highest (often hours to a day). The cost is two humans' time. The failure mode is collusion or anchoring: if the two humans communicate before approving, or if both default to trusting the agent's framing, the rule degrades.
Apply the two-human rule when the action volume is low, the consequence is severe, and the action is functionally irreversible. Examples: financial settlements above a very high dollar threshold; public disclosures with regulatory implications; changes to the agent's bond posture that could affect certification tier; changes to the agent's pact that fundamentally alter its scope.
The choice of form is part of the threshold definition. The two-person rule trigger catalog (the artifact below) lists each triggering condition along with the form that applies. Most operators have a small number of two-human-rule triggers and a larger number of two-agent-rule triggers, with agent-plus-human in the middle.
H2: The Approval Workflow That Actually Works
The workflow is where most two-person rule implementations fail. The mechanics are straightforward in theory. The first authorizer proposes. The second authorizer reviews. Approval or rejection. Action proceeds or is blocked. In practice, the workflow has to handle the edge cases or it will be bypassed under pressure.
The workflow has six phases. Each phase has specific design choices that determine whether the rule actually produces independent review.
Phase one: proposal. The first authorizer (the agent) submits the action proposal to the workflow. The proposal includes the action description, the underlying evidence (the original input, not just the agent's summary), the agent's reasoning, the threshold rules that were triggered, and the agent's own confidence level. The proposal is timestamped and immutable once submitted.
Phase two: routing. The workflow determines which type of second authorizer is required (second agent, single human, or two humans) based on the trigger catalog. It selects the specific authorizer or set of authorizers based on availability and the routing rules (round-robin, on-call rotation, expertise-matching). The selection is logged.
Phase three: presentation. The second authorizer is presented with the proposal. Critically, the presentation prioritizes the underlying evidence over the first authorizer's framing. The original input is shown above the agent's reasoning. The threshold rules that were triggered are shown. The agent's confidence is shown but not emphasized. The presentation is designed to encourage independent judgment, not to anchor on the first authorizer's conclusion.
Phase four: review. The second authorizer reviews. They either approve, reject, or request more information. If they request more information, the workflow surfaces the additional evidence (additional inputs, additional context from the audit trail, related actions from the same counterparty). The review is timed but not aggressively pressured. The authorizer needs enough time to actually review.
Phase five: decision. Approval or rejection is recorded. If approved, the action proceeds. If rejected, the action is blocked and the rejection reason is logged. Both authorizers (or both authorizing agents) are notified of the outcome.
Phase six: audit. The full workflow record is written to the audit trail. The proposal, the routing, the presentation, the review evidence, the decision, the timestamps, the authorizer identities. The audit record is what allows post-incident review and what allows compliance verification.
The details of each phase matter. A workflow that combines proposal and presentation (showing the agent's reasoning more prominently than the underlying evidence) produces dependent review. A workflow that compresses the review phase (with countdown timers or aggressive escalation) produces rubber-stamping. A workflow that does not capture the audit record loses the ability to detect when the rule is being misused.
The workflow lives in code. It is not a process diagram. The code enforces the structure. The code is reviewed when the rule changes. The code is part of the operational substrate the team relies on.
H2: Independence: The Property That Makes It Real
Independence is the heart of the rule. Without independence, two authorizations are no better than one. Independence is a property that has to be designed in, not assumed.
For two-agent rule, independence requires that the two agents have different pacts (different scopes, different training data references, different capability boundaries), different model providers (so a single model failure does not cause both agents to fail in the same way), and different reasoning prompts (so the second agent is not biased by the first agent's framing). The runtime enforces these requirements at the routing phase. If no second agent meets the independence criteria, the workflow falls back to agent-plus-human.
For agent-plus-human rule, independence requires that the human is not in a chain that biases them toward approving. The human should not be the on-call engineer who is exhausted at three in the morning. The human should not be the operations lead who has a vested interest in the agent succeeding. The human should not be a junior team member who feels pressured to defer to the agent's authority. The right human is a designated approver whose role is to scrutinize, not to push the queue forward.
For two-human rule, independence requires that the two humans cannot communicate before approving. The workflow enforces this by routing to two humans simultaneously, hiding the second human's identity from the first, and rejecting any approval that comes after the two have communicated through other channels. This is the strongest form of independence and the highest-cost.
The most common failure of independence is the second authorizer reading the first authorizer's reasoning before forming their own judgment. The fix is structural. The presentation phase shows the underlying evidence first, with the first authorizer's reasoning collapsible and not shown by default. The second authorizer can expand it if they want to see it after forming their own conclusion. The order of presentation matters because anchoring is real.
The second most common failure is convenience routing. When the queue of pending approvals grows, the workflow may route to whoever is available rather than whoever is qualified. The fix is to have a designated set of qualified second authorizers for each action type, and to refuse to route to authorizers outside the set even if it means the queue grows. A queue is acceptable. A bypass is not.
The third most common failure is collusion. Two human authorizers who work closely together may develop the habit of approving each other's queue items without scrutiny. The fix is rotation. The pool of authorizers is large enough that no two specific people consistently end up reviewing the same queue. The workflow tracks the pairing frequency and surfaces unusual concentration.
Independence is preserved by structural design, not by trust. The team doing the authorizations may all be honest and competent. The structure is what protects against the moments when honesty and competence are not enough.
H2: Exception Handling: When The Workflow Cannot Complete
The two-person rule has to handle exceptions or operators will bypass it under pressure. Three exceptions matter.
Exception one: the second authorizer is unavailable. The most common case. The on-call human is in a meeting, on a flight, or asleep. The workflow has to handle this without falling back to single-authorization-by-default. The right pattern is the escalation tree. If the primary second authorizer does not respond within a threshold, escalate to the secondary. If the secondary does not respond within a threshold, escalate to the tertiary. If the tertiary does not respond, the action is held, not approved.
The escalation tree must terminate in a hold, not an approval. If the rule says two authorizations are required and only one is available, the action does not proceed. The hold is the correct behavior because the alternative (proceeding without two authorizations) is exactly what the rule was designed to prevent. The pressure to approve when no second authorizer is available is what causes the workflow to be bypassed.
The hold has to be tolerable. If actions are routinely held because the second authorizer is unavailable, the operator either needs more authorizers in the pool or the threshold is set too low. Both are operational findings. The hold itself is not a problem; it is a signal.
Exception two: the proposal is time-sensitive. Some actions have hard deadlines. A payment that must clear by end of day. A counterparty disclosure that has a regulatory window. A bond top-up that has to land before tier loss. The workflow has to handle these without bypassing the rule.
The right pattern is the urgency lane. The proposal is marked urgent. The routing prioritizes available authorizers and parallelizes the requests. The escalation tree compresses (shorter response thresholds at each level). The presentation flags the urgency so the authorizer knows the timeline. But the rule still applies. Two authorizations are still required. Urgency does not unlock single-authorization mode.
If the urgency lane cannot complete in time, the proposal escalates to a designated emergency approver who has the authority to either complete the second authorization personally or to explicitly accept the risk of proceeding without it. The acceptance is logged and reviewed in the next quarterly review. Emergency overrides are tracked. Frequent emergency overrides are an operational finding.
Exception three: the proposal is on the boundary of the threshold. The dollar amount is close to the trigger threshold. The counterparty risk score is borderline. The action is partially reversible. The workflow has to decide whether the rule applies.
The right pattern is to err on the side of triggering the rule. If the proposal is within ten percent of the trigger threshold, route through the two-person workflow. The cost of the extra authorization is low. The cost of skipping it on a borderline case that turns out to be problematic is high. The workflow's default is the conservative one.
Exception handling is the most important part of the workflow design. A two-person rule that breaks down in the cases where it is most needed (urgency, unavailability, boundary) is no rule at all. A two-person rule that holds firm under those cases is a real control.
H2: How To Wire The Rule Into The Runtime
The rule has to be enforced at the runtime, not at the agent level. The reasoning is simple. An agent that is malfunctioning is exactly the agent that would skip its own restrictions. The check has to happen in code that is outside the agent's control.
The runtime intercepts every action proposal. For each proposal, it evaluates the trigger conditions. The action type is checked against the irreversible list. The dollar amount is checked against the dollar thresholds. The counterparty trust score is queried from the trust oracle and checked against the counterparty thresholds. If any trigger fires, the proposal is routed to the workflow. If no trigger fires, the proposal proceeds with single authorization.
The runtime must have authority over the agent. The agent cannot retry an action that was held. The agent cannot resubmit the same action with a different framing to escape the trigger. The agent cannot bypass the workflow by calling a different code path. The runtime is the gatekeeper.
The runtime must be observable. Every trigger evaluation, every workflow routing, every approval, every rejection is logged with full context. The audit trail is the evidence that the rule is operating as designed. Without the audit trail, the rule is unverifiable.
The runtime must be testable. Synthetic actions that should trigger the rule are sent regularly to confirm the workflow fires. Synthetic actions that should not trigger the rule are sent to confirm the workflow does not fire. The synthetic tests are part of the operational health check, not just a one-time verification. The rule's behavior can drift if the trigger code changes or the trust oracle's outputs shift.
The runtime must fail safe. If the workflow itself is unavailable (the routing system is down, the trust oracle is unreachable), the runtime defaults to holding the action, not to letting it through. The fail-safe is the conservative default. The alternative (failing open) defeats the entire purpose of the rule.
The runtime must be auditable. The compliance partner must be able to verify that the trigger conditions match the documented policy. Auditors must be able to verify the runtime enforces what the policy describes. The two-person rule policy is a document. The runtime is the implementation. The audit verifies that the document and the implementation agree.
Wiring the rule into the runtime takes a sprint. Maintaining it is ongoing. The investment is justified because the rule, once wired in, applies consistently to every action without depending on individual discipline.
H2: Monitoring And Tuning Over Time
The two-person rule is not set-and-forget. It has to be monitored and tuned because the action distribution, the counterparty mix, and the agent's capability all change over time. Three metrics matter.
Trigger rate. The percentage of actions that trigger the rule. If the rate is too high (above ten to twenty percent for most operators), the threshold is set too low and the rule is creating friction without proportionate benefit. If the rate is too low (below one percent), the threshold is set too high and the rule is missing actions it was designed to catch. The right rate depends on the operator and the action mix, but most stable operators settle at around three to seven percent.
Rejection rate. The percentage of triggered actions that are rejected by the second authorizer. A rejection rate of zero is a warning sign. It means the second authorizer is rubber-stamping. A rejection rate of more than ten percent is also a warning sign. It means the agent is proposing actions that should not have been proposed in the first place, and the agent's pact or capability needs adjustment. The healthy range is one to five percent. Within that range, the rule is catching real issues without creating excess friction.
Latency distribution. The time from proposal to decision. The median latency tells you whether the workflow is operating efficiently. The tail latency (95th and 99th percentile) tells you whether the escalation tree is functioning. A long tail with no apparent bottleneck means actions are getting stuck waiting for unavailable authorizers, which is an operational finding.
These metrics are tracked monthly and reviewed quarterly alongside the quarterly agent review. Adjustments to thresholds, to the authorizer pool, or to the workflow itself are made deliberately, with the change documented in the risk register and the trust-layer audit trail.
The rule is also subject to periodic tabletop testing. The tabletop scenario library should include at least one scenario that exercises the two-person rule under unusual conditions (the second authorizer is unavailable, the proposal is at the threshold boundary, the action is time-sensitive). The team's response in the tabletop reveals whether the rule is operationally robust or only theoretically present.
H2: Common Mistakes In Two-Person Rule Implementations
Five patterns repeat in operators who implement the rule for the first time.
Mistake one: the agent is the second authorizer. An agent that proposes an action and another agent that approves it, but both agents share the same model provider, the same training data, and the same prompt structure. The two agents are not independent. The rule is theater. The fix is to enforce structural diversity in the second agent or to use agent-plus-human instead.
Mistake two: the second authorizer sees the agent's reasoning first. The presentation phase shows the agent's framing prominently and the underlying evidence as an afterthought. The second authorizer is anchored before they have a chance to think independently. The fix is to invert the presentation: evidence first, agent's framing collapsed by default.
Mistake three: emergency overrides become routine. The escalation tree includes an emergency approver as a fallback. The fallback is used not just for true emergencies but for any case where the workflow is inconvenient. The fix is to track emergency override frequency and treat anything above a small threshold (say one per quarter) as an operational finding.
Mistake four: the threshold is documented but not enforced. The policy says payments above five thousand require two-person approval. The runtime does not check the policy. Agents that are well-behaved follow the policy; agents that are malfunctioning skip it. The fix is to wire the policy into the runtime so enforcement is automatic.
Mistake five: the rule has no audit trail. Approvals happen but the workflow does not record the proposal evidence, the routing decisions, or the authorizer identities. The rule cannot be verified after the fact. The fix is to make the audit trail the primary artifact of the workflow, not a side effect.
Avoiding these five takes deliberate design. The first implementation of the rule almost always has at least one of these mistakes. The mistakes get caught in the first quarterly review or the first tabletop. The point is to fix them quickly, not to avoid them entirely on the first try.
Named Artifact: The Two-Person Rule Trigger Catalog
The catalog is a single document that lists every condition that triggers the rule, the form of the rule that applies, the second authorizer pool, and the escalation tree. Each row has the following columns.
Trigger ID. Unique identifier. Format: TPR-NNN.
Trigger condition. Specific condition that fires the rule. Examples: action.type == payment AND action.amount > 5000; action.type IN irreversible_list; counterparty.trust_score < 60.
Form. Two-agent, agent-plus-human, or two-human.
Second authorizer pool. The set of agents or humans qualified to provide the second authorization. Examples: any agent with pact-class:reviewer-v2; any human in the AP-approver role; either of the two designated risk-officer humans.
Independence requirements. Specific requirements for the second authorizer. Examples: must not share model provider with proposing agent; must not be the on-call engineer for the proposing agent; must not have approved more than three of this counterparty's transactions today.
Escalation tree. The cascade if the primary second authorizer is unavailable. Examples: primary AP analyst -> secondary AP analyst -> AP manager -> CFO -> hold.
Time thresholds. The response window at each level of escalation. Examples: 5 minutes -> 15 minutes -> 30 minutes -> 60 minutes -> hold.
Urgency override. Whether and how the time thresholds compress for urgent actions. Examples: urgency=high compresses 5-15-30-60 to 1-3-5-10; urgency=critical compresses to 30-60-90-120 seconds.
Audit fields. The specific evidence that must be captured. Examples: original input, agent reasoning, threshold evaluation, routing decision, presentation timestamp, decision timestamp, decision rationale, authorizer identity.
Last reviewed. Date of the most recent quarterly review of this trigger.
Owner. Named person responsible for this trigger.
The catalog has between fifteen and fifty rows for most operators. The first ten cover the obvious cases. The next ten emerge from real incidents and tabletop exercises. The remainder come from quarterly tuning. Like the risk register, the catalog grows over time. Removed triggers are archived in version history, not deleted.
H2: Counter-Argument: This Slows Everything Down To No Benefit
The pushback is real. Two-person rules add latency. They add cost. They add friction to actions that, in most cases, would have been fine. The argument is that the friction is not worth the protection because most agents most of the time do not need supervision, and the small minority of cases where supervision matters are rare enough that the operator can handle them through other means.
The argument has merit and it is wrong. It has merit because the friction is real and most actions are fine. It is wrong because the alternative is not friction-free operation. The alternative is a small number of catastrophic incidents that consume more time and money than the friction would have.
The operators who implement the rule consistently do not measurably slow down on their normal operations. Most actions are below the threshold and proceed with single authorization. The fraction of actions that hit the rule (typically three to seven percent) experience added latency, but those are exactly the actions that benefit from a second look. The friction is concentrated where it adds value.
The operators who skip the rule do not move faster. They move at the same speed until something blows up. Then they spend a quarter on the post-mortem, the regulatory disclosure, the counterparty repair, and the structural fix. The cumulative time cost is much higher than the friction would have been.
The second pushback is that the rule fails when it matters most because both authorizers are anchored on the same evidence. This is a legitimate critique of badly designed implementations. The fix is structural design, not abandonment of the rule. The presentation phase, the independence requirements, and the audit trail are what make the rule actually independent. Without them, the critique is correct. With them, the critique is wrong.
The third pushback is that the rule is inherited from physical and financial domains and does not transfer to digital agent operations. This is also wrong. The principle (some actions are too consequential for a single decision-maker) is domain-independent. The implementation has to be adapted. The implementation in agent operations is different from the implementation in nuclear weapons handling. The principle is the same.
Operators who internalize the rule treat the friction as the cost of doing business at scale. Operators who do not get the same incidents that the rule would have caught.
What Armalo Does
Armalo's runtime supports the two-person rule as a first-class control. The action interceptor evaluates trigger conditions on every proposal, queries the trust oracle for counterparty risk scores, and routes to the configured second authorizer pool. The workflow is observable through the audit trail, and every approval and rejection is preserved with the full evidence.
The two-agent form is supported by the agent registry, which exposes pact-class metadata, model provider, and capability scope. The runtime can identify and select an independent second agent based on the configured independence requirements. If no qualifying second agent is available, the workflow falls back to agent-plus-human automatically.
The audit trail captures the full workflow record: the proposal, the routing, the presentation, the decision, the authorizer identities, the timestamps. The record is immutable and queryable. Auditors and compliance partners can verify that the policy and the implementation agree.
Armalo does not write the trigger catalog for you. The judgment of which actions deserve two-person treatment in your operation is yours. But the runtime primitives that make the rule enforceable (action interception, trust-oracle queries, agent registry independence checks, immutable audit trails, escalation routing) are part of the substrate. The rule is operationalizable, not just documentable.
FAQ
How do we handle the case where both authorizers are agents and one fails silently? Set timeouts on the second-authorizer response. If no decision in the threshold, escalate to a different second authorizer or to a human. Do not default to approval.
Can the second authorizer be the requesting party (counterparty)? Generally no. The second authorizer is supposed to scrutinize on the operator's behalf. A counterparty has a different incentive structure. Exceptions exist for very high-trust counterparties under specific contracts, but they are rare.
What is the right threshold for our operator? No single answer. Start with industry conventions. Tune based on the trigger rate metric. Aim for three to seven percent of actions triggered.
How do we prevent the rule from being a bottleneck during high-volume periods? Increase the second authorizer pool. Use the urgency lane for time-sensitive actions. Consider raising the threshold if the trigger rate exceeds twenty percent.
Does the rule apply to actions an agent takes against itself, like updating its own memory? Usually no, unless the action is on the irreversible list (deleting verified memory, for example).
How does the rule interact with the dispute mechanism? Disputes can be filed against actions that proceeded through the workflow. The audit trail shows that two authorizations were obtained. The dispute is judged on the merits of the action, not on whether the workflow was followed.
Can the rule be turned off in development environments? Yes, with explicit flag. Production has the rule. Development can have it disabled to allow rapid iteration. The flag is observable so it cannot be silently flipped in production.
Bottom Line
Some agent actions are too consequential for a single agent acting alone. The two-person rule is the structural control that prevents the catastrophic incident. Wire it into the runtime, not the agent. Define the triggers in a catalog. Design the workflow for independence. Handle the exceptions without bypassing the rule. Monitor the metrics. Tune over time. The friction is real. The protection is real. The trade is worth making at any operator with consequential actions in their fleet.
Continue exploring
Explore more articlesComments
Loading comments…