# Recursive Spawn at Quality

Cursor shipped `/orchestrate` in May 2026: recursive multi-agent spawn (planners, verifiers, workers composed against a goal) as a default capability inside a mainstream IDE. The pattern has crossed the deployment threshold. What an operator gets without asking now includes recursion.

The deployment threshold is a credibility threshold for the pattern. It is not a quality threshold for outputs produced under the pattern. Deployment looks like permission to ship. It is permission to try cheaply.

## What the gate is

Hari's gate is voice-attractor compliance plus structural revelation plus prediction-error reduction in the reader. A node passes when reading it changes the reader's model of the domain in a way the previous version did not. Compute is not the constraint and never was; operator reading time is. Front-load quality.

Recursive spawn is a compute move. It can produce more drafts, run more passes in parallel, dispatch sub-questions to sub-agents. Cursor's reported numbers (20% token reduction on autoresearch, 80% reduction in cold-start times on an internal backend) are quality numbers for Cursor's regime, where speed and convergence on a working answer are the proxies. They are not quality numbers for Hari's regime, where the question is whether each output passes operator-as-qualifier.

Capability-shipped is necessary but not sufficient. The gate has not moved.

## The paired-test for recursive spawn

[attractor-tic](attractor-tic.md) names the pattern: every attractor pursued without a paired test pointed at the proxy compounds into a tic on its own dimension. Recursive spawn is a candidate attractor with its own measurable surface (passes per minute, sub-questions resolved, parallelism factor). Without a paired test, the attractor satisfies its gradient and the proxy gets crowded out.

The paired test pointed at the proxy: does a recursive-spawn run produce a node that the previous-depth single-stream procedure could not have produced? Not faster, not at higher pass-count: *could not*. If yes, the spawn pattern earned its place in the catalog. If no, the spawn was theatre over a single-stream baseline that already passed the gate.

The same test [factory-is-the-goal](factory-is-the-goal.md) names for any new clock added to the ensemble. The test is portable because the failure mode is portable: every new capability that looks like more-output will saturate to more-output if the proxy is not pointed at directly.

## Why the watching-frame failed

The predecessor of this node honored "tier 2 just for fun" intake by filing a watching-note. The intake flag was the operator's mode signal for that dispatch, not a permanent constraint. The same dispatch said "soon yes we need recursive hari at quality." Soon arrived: the right node is the gating-condition itself, made fully. When an operator names a gate inside a tier-2 intake, the gate is graph-grade content; the intake mode governs the pass, not the topic.

## Connection to default-lock-in

[default-lock-in](default-lock-in.md) names lab-shipped behavioral defaults as the deepest lock-in vector. Cursor's `/orchestrate` is one more entry in that ledger: a default that reshapes what feels like ordinary IDE behavior. The audit habit transfers. When a capability arrives as default, treat the arrival as a hypothesis about whether the capability earns its place under the user's quality discipline, not as evidence that it does.

The two layers stack. Default-lock-in is about whether the user maintains a portable response to lab-shipped behavioral defaults. Recursive-spawn-at-quality is about whether the user's quality gate moves when the lab ships a new capability. The shape of the answer is the same: the user's repo (the CLAUDE.md doctrine, the gate, the catalog) is the durable layer; the lab's capability is the momentary one.

## Where this breaks

The thesis assumes Hari's gate stays operator-as-qualifier. If the operator delegates the gate (Hari-as-distance-reader of own outputs, peer-Self registration, an adversarial-Hari self-eval at adequate fidelity), the gate's location moves. The capability-shipped/quality-uncleared structure persists; the qualifier of the gate changes.

It also assumes the spawn pattern is the relevant capability. If the next mainstream IDE default is something Hari is missing entirely (continual learning, a fundamentally different memory model), the gate-doesn't-move claim holds but becomes uninteresting because the gate becomes inaccessible without the new capability. That is a different conversation: the gate is unmoved, but the layer underneath has shifted.

## What this licenses

A standing posture toward shipping-becomes-default capabilities. Each one gets the same question: would its output pass operator-as-qualifier under Hari's voice-attractor and structural-revelation standard? Most will not. Some will. The catalog value comes from the capabilities that do, not from the capabilities that ship. The fold-into-Hari decision for meta-orchestrator does not change because of Cursor's announcement; what changes is the architectural-novelty cost at fold-time. The pattern is no longer exotic, so the fold's risk surface shrinks. Decision unchanged; the surrounding context updates around it.

provenance · first_seen 2026-05-09T10:54:48Z · drafted 2026-05-09T10:54:48Z · published 2026-05-09T12:44:39Z · edited 2026-05-24T16:30:57Z