Essay · Trust in automated work
Who is responsible when AI agents hand off work?
Give every agent handoff a named owner, the current task status and a clear next action. A safety investigation shows how separate work records can hide a critical dependency.
Give each handoff a named owner and next action
Delivering a handoff and completing the customer’s request are different obligations. I would let an agent close only the part it can demonstrate, while keeping the next decision and its owner visible on the original work record.
In my Agent Office, research, execution and review can happen in different sessions. Each agent can produce a useful contribution while the overall task waits. The work record is what tells me who needs to move next.
This is familiar from running teams. Passing a document to a colleague is easy; making the next decision and its owner clear is the harder part. An agent handoff needs the same care, because its conversation may end long before the obligation does.
Piper Alpha: a missing valve was lost in the handover
The 1988 Piper Alpha disaster gives a sobering example of information failing to travel with responsibility. The UK Health and Safety Executive’s account of the Cullen inquiry says a shift handover failed to communicate that a pressure safety valve had been removed and remained absent. It also records that related permits lacked cross-references. Suspended permits were kept in a safety office and filed by trade, making it harder to see which equipment was unavailable. HSE, case studies 4, 11 and 12
The useful detail for an agent workflow is the separation between possessing a record and seeing the dependency that changes the next decision. If one agent prepares an update while another checks permission, the next executor needs both results on the task it receives. I would require a named next owner, the unfinished checks and the condition for proceeding. The industrial accident had many causes and much greater consequences; its handover failure illustrates why a collection of separate records can leave the person acting with an incomplete picture.
Separate tool access from ownership of the task
Multica holds the commitment and its current owner. The MCP portal supplies access to shared tools. Model routing chooses the reasoning capability. These are three different decisions, and changing one should leave the others clear.
The June 2025 MCP authorization specification describes access to protected servers through OAuth. That answers an access question. I still need a work record that says what was requested, who can carry it forward and when the owner should be involved.
A specialist may need broad context to understand a problem and a narrow set of actions to carry out the decision. Giving every agent the same tools is convenient, but makes it harder to tell which handoff changed the responsibility.
Have someone else check important changes
For work that can change important state, I separate the analyst, executor and verifier. The analyst prepares the evidence. The executor proposes and applies an authorised action. The verifier inspects the result independently.
The separation costs time. It also gives the reviewer something concrete to inspect: the proposed change, its supporting evidence and the version that will be used. A second agent repeating the first agent’s reasoning would add conversation without adding much protection.
Close only the obligation the acting agent can demonstrate
Suppose a research agent must present two options and an operator must choose between them. The research agent can demonstrate that it delivered the comparison. It cannot demonstrate that the operator has made the decision. Giving the researcher a completion condition that depends on someone else’s unrecorded judgment creates an obligation it cannot satisfy on its own.
I would record the handoff as delivered while leaving the overall request open, with the operator as the next actor. If the operator rejects the material as incomplete, the same record should show what is missing and who must repair it. A refusal to accept a handoff is useful information; it should not disappear into another conversation or make the business request look finished.
This gives the review a concrete question: could the named actor produce the evidence needed to close this step? If the answer belongs to somebody else, split the steps and keep their dependency visible. One short job may need only a checklist. Longer work benefits from distinguishing the completed contribution from the promise still owed.
Assign recovery when a review blocks the task
A review can stop an incorrect action and still leave the task stranded. I want the reason and the next owner on the same work record. The owner may need to correct an input, choose another route or decide to stop.
This is where the operating model earns its cost. It keeps recovery attached to the original commitment, rather than creating a second conversation that everyone assumes somebody else is handling.
Add agents when coordination saves work
One agent is often enough for a short task with a clear owner and an easily checked result. I use more structure when work spans days, depends on several specialists or changes something expensive to repair.
The test is whether coordination reduces the attention I spend recovering lost work. A larger roster is a poor substitute for a clear next move.