For a cross-company handoff, the practical check is simple: before any retry, ask whether the append-only trail already records acknowledgment or terminal failure. If it does not, the boundary evidence says ownership is still unsettled, and another attempt only multiplies ambiguity.

Options compared: direct, relay, shared queue
When teams compare handoff patterns, the practical question is not which one moves a payload fastest. It is which one can be audited as a single handoff, with one accountable owner on each side and one shared record of what happened. On that test, three patterns show up: direct agent-to-agent calls, a manual ticket bridge, and a relay-owned handoff. Only one of them is suitable as the default control.
Direct A-to-B calls look clean in a diagram, but they fail the single-audit-trail test when each company keeps its own log. Company A can record “sent,” while Company B records “received,” and neither log is authoritative over the other. That creates two truths and no single owner table a reviewer can use to settle state. If the two logs disagree, the system has no shared place to show whether the handoff was acknowledged, rejected, or still pending. A direct call may still carry business data, but as a control pattern it is weak because the evidence is split across company boundaries.
A manual ticket bridge passes the audit test more easily because the ticket can become the shared record. The problem is latency and retry behavior. A human has to notice the ticket, interpret it, and act on it, which slows the handoff. Manual processes also tend to permit manual retries: someone resubmits the ticket, forwards it again, or opens a duplicate when the first one seems stuck. That makes the ticket useful as a fallback when no relay is available, but not as the preferred pattern for agent-to-agent work where state must be controlled before payload movement.
The winner is the relay-owned handoff. In this comparison, it is the only pattern that can enforce two named owners and one state machine before the payload crosses the boundary. The relay gives both companies a common control point: one append-only record for the handoff, one place to check state, and one place where acknowledgment or terminal failure is written. That structure also makes blind retries suppressible, because the system can check the record before taking any further action.
The comparison can be summarized this way:
| Pattern | Audit result | Retry control | Verdict |
|---|---|---|---|
| Direct A-to-B calls with separate logs | Two logs create two truths | No shared record to gate retries | Reject as the control pattern |
| Manual ticket bridge | Can provide a shared record | Human latency; manual retries are common | Fallback only when no relay exists |
| Relay-owned handoff | One append-only record | Blind retries can be suppressed | Winner |
This is why the relay pattern should be treated as the default for cross-company agent handoffs. It does not rely on each side trusting the other’s private log, and it does not depend on a person to keep the shared record coherent. Most importantly, it puts control ahead of transport: the handoff is governed before the payload crosses, rather than reconstructed afterward from scattered evidence.

Costs and numbers that matter
This section fixes the counts that the rest of the rule depends on: 2 owners, 1 audit trail, and 0 blind retries. I treat those as method counts, not aspirations. Before a cross-company agent handoff is allowed to move, the record must show exactly two named accountable owners, one on each side, and exactly one append-only audit trail keyed to that handoff. If any of those counts is wrong, the handoff is not ready.
The owner count is exactly 2 named accountable owners: one per company. Zero owners means stop, because no one is accountable for the state of the handoff on either side. Three or more owners means the handoff has become a committee, and I require it to be split into separate handoff records with one accountable owner per company on each record. A shared distribution list, a department alias, or a rotating on-call label does not satisfy the count; the record must name the person who owns the handoff outcome for that company.
The audit count is exactly 1 append-only record per handoff. Every state change must append to that record, including acknowledgment, terminal failure, and any correction after a defect is found. If a second system of record starts carrying handoff state, I treat that as a defect to reconcile before any retry is considered. The check is simple: pick the handoff ID, read the trail, and confirm that the trail alone explains the current state. If two systems disagree, the handoff is not eligible for another attempt.
The retry count is 0 blind retries. A retry is blind when it fires before the audit trail records either acknowledgment or terminal failure. The cost of a blind retry is the duplicate action’s value multiplied by the retry count. If the duplicate action’s value is unknown, I do not assign it a convenient estimate; I suppress the retry until the trail supplies a recorded outcome. That keeps the arithmetic honest: unknown value times an unapproved retry count is not a business case, it is an unpriced exposure.
These counts also fit how I read boundary rules in the source material.
The US EPA page on determining organizational boundaries, dated Feb 10, 2026, points to the GHG Protocol’s methods for setting boundaries, and GreenCalculus’s glossary (Nov 6, 2026) explains that organizational boundaries decide which operations a company counts as its own.
That boundary logic supports the owner count here: each company must name its own accountable owner before the handoff crosses the line. The Wiley review of organizational boundary definitions, functions, and properties likewise treats boundaries as structural features of organizations, not informal gestures.I use the counts as a preflight table. It is not a scoring model; it is a gate.
| Count | Required value | Check before handoff proceeds |
|---|---|---|
| Owners | 2 | One named accountable owner per company; 0 means stop, 3 or more means split into separate handoff records. |
| Audit trails | 1 | One append-only record per handoff ID; every state change appends; a second system of record is reconciled first. |
| Blind retries | 0 | No automatic retry until the trail records acknowledgment or terminal failure; duplicate cost is value times retry count. |

What the evidence does NOT establish
This section is the limit check on the 2-1-0 rule, not a performance claim for it.
The grounding sources here establish boundary concepts, not measured cross-company agent behavior. They contain no measured cross-company agent latency, no duplicate-work rate, and no retry-amplification figure. Any timeout, breaker, or retry-budget value imported from other agent-failure material should be treated as unverified for this handoff rule; before relying on such a number, check whether the named source actually measures cross-company handoff behavior rather than a different failure surface.
The check to carry forward is the 2-1-0 gate as defined in the counts section above.
The edge cases matter because legal structure changes who can own a handoff. The sources support that much: the WRI PDF on organizational boundaries notes that companies operate through incorporated and non-incorporated joint ventures, subsidiaries, and other structures, while the ASUENE manufacturer’s guide addresses joint venture treatment and contract manufacturer rules.
GreenCalculus (Nov 6, 2026) describes equity share, financial control, and operational control as different ways to decide which operations a company counts as its own.
That is useful for assigning accountability, but it is not proof that an agent handoff will perform well under any of those structures.| Edge case | What the 2-1-0 rule does | Reader check |
|---|---|---|
| Shared JV or contract-manufacturer control | Rule breaks unless a relay custodian is named. | Confirm that control shared across entities does not leave the handoff ID without one accountable writer on each side. |
| Single parent-subsidiary | Rule wins because one owner per legal entity is clear. | Name one owner for the parent and one for the subsidiary; do not collapse them into one informal owner. |
| Vendor outage | Rule wins if zero blind retries are allowed and terminal failure is recorded. | Suppress retries until the append-only trail records acknowledgment or terminal failure. |
The shared JV or contract-manufacturer case is the weak point. Where control is shared, the boundary question becomes who may speak for the operation, not merely which company has an ownership percentage or a contract. If no relay custodian is named, the rule breaks at the ownership step: there is no clean pair of accountable owners and no reliable path for the handoff ID to be written through. The practical check is to name that custodian before the handoff begins, then keep the same one-record-per-ID discipline.
The parent-subsidiary case is cleaner. A single legal entity boundary does not erase the need for two owners; it makes them easier to identify. One owner belongs to the parent, one to the subsidiary, and the handoff still needs its single append-only record. The vendor-outage case is also favorable, but only under the zero-retry condition. If the trail records terminal failure, the handoff has a defensible end state. If the system retries blindly while the record is silent, the rule loses its control point.
When the owner assignment is ambiguous, do not treat the boundary source as a substitute for naming people. GreenCalculus shows why the choice of equity share, financial control, or operational control can change what a company counts; the EPA page likewise frames organizational boundaries as defined methods under the GHG Protocol. Those sources support careful ownership analysis. They do not supply a measured safety threshold for agent retries, duplicate work, or latency. The safe reading is narrower: resolve ownership first, write one audit record per handoff ID, and keep retries suppressed until that record says the handoff was acknowledged or ended in terminal failure.

Invoice agent handoff
In this illustrative scenario, Company A’s support agent sends invoice dispute H-1042 to Company B’s billing agent for a $180 duplicate charge with a 30-minute service-level agreement (SLA). Company A names R. Patel as its accountable owner, Company B names L. Chen as its accountable owner, the retry budget is set to 0, and both parties write to one append-only record for H-1042. These inputs are not measured production data; they provide a concrete case for checking how ownership, recording, and retry suppression work together during a live dispute.
| Time | Actor | Appended entry | Audit state |
|---|---|---|---|
| t+0 min | Company A support agent | Dispute H-1042 transmitted | sent |
| t+4 min | L. Chen (Company B) | Acknowledgment recorded | ack |
| t+22 min | Company B billing agent | needs-manual-review returned |
terminal |
The timeline shows how the record governs action. L. Chen’s acknowledgment at t+4 minutes confirms receipt within the 30-minute SLA window, and the terminal needs-manual-review entry at t+22 minutes stops automated handling with a definitive outcome. When R. Patel issues the manual credit at t+47 minutes, the resolution lands 17 minutes past the SLA (47 − 30 = 17). Across the handoff, 0 automatic retries fire and the audit retains exactly 3 state entries: sent, ack, and terminal.
The wrong-way cost becomes clear at t+22 minutes. The disputed amount is $180. If a blind retry fired after the terminal entry, Company B could issue a second credit against the same dispute, producing $180 (intended credit) + $180 (duplicate credit) = $360 in total exposure. Suppressing retries until the record shows acknowledgment or terminal failure prevents that duplicate payout while limiting the financial outcome to the single manual credit R. Patel authorizes.
Before a similar invoice handoff, the operating check is concrete: confirm R. Patel’s and L. Chen’s counterparts are named on each side, confirm one append-only record is open for the dispute, and confirm automated retries remain suppressed until that record contains an acknowledgment or a terminal outcome. In H-1042, those controls did not eliminate the SLA miss, but they did contain the financial risk to a single $180 credit and left a complete, auditable sequence of decisions.

Rules: two owners, one trail, zero retries
This section gives the if/then rules that enforce the canonical rule: before any cross-company agent handoff, require exactly two named accountable owners, exactly one append-only audit record per handoff ID, and zero automatic retries until that record shows acknowledgment or terminal failure. Treat these as gate conditions, not guidance. If a gate fails, the handoff does not proceed.
If either company cannot name exactly one accountable owner, then do not create the handoff ID. “The integration team,” “the vendor contact,” or “whoever is on call” is not a named owner. The check is simple: each company must be able to point to one person who owns the handoff outcome on its side, including escalation and acceptance of the recorded result. If one side has two owners, no owner, or only a role title, stop before any identifier is issued. This prevents ambiguity from being encoded into the handoff itself.
If the handoff would write to two systems of record, then stop and consolidate to one append-only audit trail before sending payload. The test is not where the payload travels; it is where the authoritative handoff status is written. If one system records “sent,” another records “received,” and a third records “approved,” the companies no longer share a single account of what happened. Consolidate those status writes into one append-only trail keyed by the handoff ID, then allow the payload to move. The trail is the control surface; other systems may read from it, but they should not compete with it as the record of handoff state.
If a retry is requested before the audit record shows acknowledgment or terminal failure, then reject it as a blind retry and log the rejection. The audit record must show one of two closing states before any retry logic is permitted: acknowledgment by the receiving owner, or terminal failure recorded in the same trail. Until then, the correct response to timeout anxiety, queue delay, or an unanswered message is not another send. It is a check of the trail. If the trail is silent, the handoff remains unresolved, and the retry request itself becomes evidence that the control gate is working.
These rules also clarify where organizational boundaries matter. The US EPA’s “Determine Organizational Boundaries” page notes that the GHG Protocol defines methods for setting boundaries in its Corporate Accounting and Reporting Standard and Corporate Value Chain standard, and GreenCalculus’s organizational-boundary glossary, updated Nov 15 2026, describes equity-share and control approaches. The agent-handoff parallel is limited but useful: before counting or transferring anything across a boundary, the responsible party must be fixed. In this rule set, that means one owner per company, one append-only trail per handoff ID, and no retry until the trail closes the state.
Use the gates in order. First, confirm two named owners. Second, confirm one append-only audit trail. Third, confirm the trail shows acknowledgment or terminal failure before allowing any retry. A handoff that satisfies all three is controlled; a handoff that misses any one of them is incomplete, even if a payload appears to have moved.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Before any cross-company agent handoff, name exactly two accountable owners: one per company, with the sender and receiver explicitly identified. | This enforces the reader rule that no handoff is sent unless exactly two owners are in place. |
| 2 | Create one append-only audit record keyed to the handoff ID, and require both named owners to write only to that same record. | This preserves exactly one audit trail per handoff ID across the organizational boundary. |
| 3 | Confirm the audit record shows receiver acknowledgment or terminal failure before allowing any retry behavior. | This keeps automatic retries at zero until the record proves the handoff reached a defined end state. |
| 4 | If either company has not named its owner, stop the handoff before send. | A missing owner breaks the canonical decision rule, regardless of company policies and practices. |
| 5 | If the handoff ID has no append-only record, or has more than one record, stop the handoff before send. | The audit trail must remain singular and append-only for every cross-company handoff. |
| 6 | When reviewing Discovery|PatSnap’s point that barriers and actions are predicated to a large degree on boundaries controlled by organizational policies and practices, document those boundaries in the same handoff record rather than routing around them. | This keeps the handoff controlled at the company boundary while preserving the single audit trail. |
Frequently Asked Questions
What must be checked before retrying a cross-company handoff?
Before any retry, ask whether the append-only trail already records acknowledgment or terminal failure.
What does it mean if the append-only trail lacks acknowledgment or terminal failure?
The boundary evidence says ownership is still unsettled, and another attempt only multiplies ambiguity.
Which three handoff patterns are compared?
Three patterns show up: direct agent-to-agent calls, a manual ticket bridge, and a relay-owned handoff.
Why do direct A-to-B calls fail the single-audit-trail test?
They fail the single-audit-trail test when each company keeps its own log.
What conflicting records can arise in direct calls?
Company A can record “sent,” while Company B records “received,” and neither log is authoritative over the other.
What happens if the two company logs disagree?
If the two logs disagree, the system has no shared place to show whether the handoff was acknowledged, rejected, or still pending.
Quick answers
| What should be checked before retrying a cross-company handoff? | Before any retry, ask whether the append-only trail already records acknowledgment or terminal failure. |
| What happens if the append-only trail does not record acknowledgment or terminal failure? | The boundary evidence says ownership is still unsettled, and another attempt only multiplies ambiguity. |
| What is the practical question when comparing handoff patterns? | The practical question is which one can be audited as a single handoff, with one accountable owner on each side and one shared record of what happened. |
| Which three handoff patterns are compared in the article? | The three patterns are direct agent-to-agent calls, a manual ticket bridge, and a relay-owned handoff. |
| Why do direct agent-to-agent calls fail the single-audit-trail test? | They fail because each company keeps its own log, creating two truths and no single owner table a reviewer can use to settle state. |
Also worth reading: Secure AI agent handoffs without leaking context: Secure AI agent handoffs without · Multi agent pipeline errors 2026: schema handoffs cut 40% vs free text: Multi agent pipeline errors 2026: · Multi-agent workflow handoffs 2026: 3-step handshake protocol for reliability: Multi-agent workflow handoffs 2026: 3-step