// MACHINE-WORKER MANAGEMENT · TRUST DENSITY · June 20, 2026

Machine-Worker Management Needs Custody Chains Before Agents Hand Off Sensitive Work

When autonomous workers pass sensitive work to each other without custody, ownership evaporates exactly where risk concentrates.

WhatTraceable handoff

A custody chain is the operating record that proves which worker held sensitive work at each moment and under which authority.

WhyRisk travels

Autonomous labor does not fail only at task execution. It fails when context, authority, and accountability are stripped during routing.

ConsequenceSafe delegation

Operators get faster machine collaboration without turning every cross-agent handoff into a hidden liability transfer.

Machine-worker handoffs are where clean automation becomes messy management.

Single-agent demos make work look tidy: one prompt, one tool call, one output. Real companies do not work that way. A growth worker finds a customer objection, a RevOps worker updates the account, a support worker drafts a response, a finance worker checks a discount, and a manager agent decides whether to escalate.

The danger is not that one worker touches the task. The danger is that the task changes hands until nobody can explain which worker owned the decision, which evidence was trusted, and which constraints were still active.

That is why custody chains matter. They make delegation visible before autonomous workers begin treating sensitive work like ordinary queue traffic.

Four handoffs that need custody.

01Customer promise

When an agent passes a pricing, delivery, support, or product claim to another worker, the receiving worker must inherit the exact approved promise boundary.

02Data access

When work requires customer, employee, financial, or partner data, the handoff must show which fields are needed, why, and when access expires.

03Revenue exception

When an agent routes a discount, refund, renewal risk, or credit change, custody should include authority level, evidence, threshold, and approver.

04Human escalation

When a worker asks for human judgment, the chain should preserve what the machine already did, what it refused to do, and what decision is actually needed.

The worst handoff is the one that looks like progress.

Imagine a growth agent spots a strategic account asking for an implementation timeline. It routes the thread to a customer-success worker. That worker asks a product-planning worker for availability. The product worker infers a likely ship window from roadmap notes and sends it back. The customer-success worker drafts a confident reply.

Everyone can say they only performed their local task. Nobody can prove that the final promise was authorized. The context lost its warning labels while moving through the machine workforce.

That is the core failure: handoffs can launder uncertainty, expand authority, and turn partial evidence into customer-facing commitments.

The custody chain receipt

Sender

The worker, rule, or human that transferred the task, including the reason the work left its original lane.

Receiver

The worker allowed to act next, the exact mandate it receives, and any actions explicitly forbidden.

Evidence packet

The source records, customer messages, internal notes, policy links, and confidence labels that must travel with the task.

Scope limit

The time limit, data boundary, approval threshold, affected customer segment, and rollback condition attached to the handoff.

Return path

The condition that sends the task back to the sender, escalates to a human, or blocks further autonomous action.

Custody is not surveillance. It is how autonomous teams keep promises attached to owners.

The point is not to create a giant audit museum. The point is to make every sensitive handoff carry enough context for the next worker to act safely. Low-risk work can flow quickly. High-risk work can still move, but only with the receipt that preserves authority.

Good custody chains should be short, automatic, and hard to bypass. They should appear beside the work, not in a forgotten compliance export. If the receiving worker cannot display the chain, it should not act.

A machine worker should never receive more authority than the handoff proves.

Install one custody chain this week.

PickOne route

Choose a recurring cross-agent handoff: growth to RevOps, support to product, sales to finance, or agent to human manager.

DefineOne boundary

Name the sensitive item that must not be lost: promise, data access, discount authority, customer segment, or escalation reason.

RequireOne receipt

Make the handoff include sender, receiver, evidence, scope, expiry, forbidden actions, and return path.

BlockOne failure

If the custody chain is missing or stale, stop the receiving worker from acting and return the task to the owner.

Concrete takeaway for this week

Before expanding multi-agent routing, put a custody chain on one sensitive handoff.

The blunt rule: if sensitive work changes hands, custody must change with it.