Delegation and Escalation Paths
Chapter 3 put permission boundaries on a single agent. In a team, the permission question gains a dimension: each agent needs an explicit escalation path — what it may decide on its own, and what it must send up.
Without One, the Team Goes to Extremes
Take the escalation path away and the team drifts toward one of two extremes.
One extreme is escalating everything. Every agent pushes anything uncertain upward, all decisions pile up at the next level or at the human, and the team has plenty of nodes but its throughput is jammed at a bottleneck. Parallelism loses its point. The other extreme is carrying everything alone. Every agent assumes the call is its own to make, so overreach and irreversible actions happen in every corner, and disaster is only a matter of time.
Both extremes share one root: nobody said what this role should finish itself and what it should escalate.
An Escalation Path Defines When to Hand It Up
A clear escalation path answers three questions: which decisions this role finishes itself, which go to a higher-tier agent or a human, and what triggers the handover.
The triggers are usually these: uncertainty above a threshold (the agent itself is not confident), an irreversible external action (a mistake cannot be recalled), or something outside the role's remit (not its call to make). Hit any one of them and the agent should not tough it out — it should escalate along the path. This generalizes chapter 3's rule of allocating delegation by reversibility from one agent to an entire chain: what is reversible, certain, and within the role's remit, the role finishes; what is irreversible, uncertain, or out of bounds goes up.
An Escalation Path Gives Risk Somewhere to Go
With one, the team can run free and stay contained at the same time.
Free, because the mass of in-remit, reversible decisions no longer needs layer-by-layer sign-off; every role advances in parallel. Contained, because the high-risk nodes — irreversible actions, high uncertainty — have a defined destination and travel the path to somewhere that can bear the responsibility. A team's autonomy equals clear delegation plus an explicit escalation path; without the latter, autonomy has only two endings, runaway or paralysis.
The path often ends at a human. The human is not an exception in this picture — the human is a role too, with responsibilities, permissions, and deliverables: which class of decisions to make, up to which level to approve. What reaches the human should be conclusions, evidence, and risk flags, not the raw stream. Fail to design the human as a role and escalation becomes dumping: everything floods toward one infinitely absorbent inbox, the human drowns, and the escalation path stops working. A mixed team of humans and agents keeps running for one reason: both sides are roles, and the interface constraints bind both sides equally.