Delegation requires
Ceilings
Every helper makes a reasonable request. Together, they overwhelm the system they came to investigate.
Part of Trustworthy Autonomy ↗Chapters in this collection 07 / 09
Delegation gives another agent a bounded task. A ceiling limits what the whole delegation can consume or change, even when every individual request seems reasonable.
Imagine fifty infrastructure agents notice the same regional slowdown. Each launches four helpers. Each helper makes a query and retries it twice.
That is six hundred query attempts arriving at an observability service already having a bad afternoon. The assistants have become part of the incident.
Small decisions can add up to a large one#
No helper intended to overload anything. Each saw a local task and a modest retry allowance. The missing perspective was the total.
This is a collective-action problem: individually sensible choices can produce a shared result nobody would choose. Coordinating the whole matters as much as improving each participant.
Agent systems need both local boundaries and aggregate limits. A child should receive a narrower task, data scope, tool set, deadline, and budget. The system also needs to account for the parent's entire tree and the fleet using the same dependencies.
A parent with one hundred queries remaining cannot create ten children and give each a fresh allowance of one hundred.
Responsibility follows the work#
Delegating an investigation does not transfer away accountability for its cost or completeness. A parent must know what each child was asked to do, when it must stop, and how to represent partial results.
Silence is especially dangerous. A helper that times out has supplied missing evidence. Its absence cannot be quietly counted as agreement.
The helpers meet a shared budget
Illustrative limits, not universal defaults. Compare local permission with fleet impact.
- Reviewers × helpers
- 50 × 4
- Attempts each
- 3
Everyone stays inside their own limit
- What happens
- Two hundred helpers each make an initial query and two retries: six hundred possible attempts.
- What it means
- A local retry limit does not bound the combined pressure on the dependency.
- What the system does
- Count attempts against shared budgets, including retries and descendants.
Inspect the example record
{
"reviewers": 50,
"helpers_per_reviewer": 4,
"attempts_per_helper": 3,
"total_attempts": 600
}- Fleet query budget
- 120 / minute
- Concurrent queries
- 8 maximum
The tree cannot mint more budget
- What happens
- The scheduler enforces a shared attempt budget and concurrency cap. Child reservations consume the parent's remaining allowance.
- What it means
- Launching more helpers changes how work is divided, not how much authority or capacity exists.
- What the system does
- Queue, reduce scope, or stop work when limits are reached. Reserve capacity for essential checks.
Inspect the example record
{
"parent_remaining_queries": 20,
"child_reserved_queries": 5,
"parent_unreserved_queries": 15,
"fleet_attempts_per_minute": 120,
"max_concurrent": 8
}- Helper status
- Timed out
- Evidence status
- Missing
An incomplete result stays incomplete
- What happens
- A helper finishes only two of three regional comparisons before its deadline.
- What it means
- The parent has partial coverage and must keep that gap visible in its verdict.
- What the system does
- Return the completed evidence, missing region, and termination reason; do not infer success.
Inspect the example record
{
"completed_regions": [
"A",
"B"
],
"missing_regions": [
"C"
],
"termination": "deadline",
"parent_effect": "coverage gap; no all-region conclusion"
}A good delegation can fit on a card#
State the task, permitted tools, data boundary, query and retry budget, deadline, termination condition, and required result format. A child should be able to understand its assignment without hidden shared context.
Enforce maximum depth and fan-out. Share cancellation. When the parent stops, its descendants should not continue spending or acting as if nothing changed.
Keep fleet limits outside individual model sessions. A limit remembered only in a prompt can disappear when work is split, retried, or resumed elsewhere.
The practical test is multiplicative: increase reviewers, helpers, retries, and nesting together. Total exposure should remain inside the system's declared ceiling, and unfinished work should remain visible.
Delegation can multiply attention. It must not multiply permission by accident.
Further reading: Google SRE's Handling Overload discusses retry budgets and protecting shared capacity.