authority-delegation-garden-patch

Orchestrator-Worker Separation in Personal Multi-Agent Systems

Heart

Domain owners on top, task executors below. A small number of heavyweight orchestrators hold persistent state and coordinate work within their domains; a larger pool of lightweight workers execute bounded commissions without accumulating identity across invocations. This two-tier split resolves three ceilings simultaneously — context, attention, and cost — that a flat network or monolithic agent cannot.

Problem

Personal multi-agent systems face context, attention, and cost ceilings that a flat peer network cannot solve — no single agent holds enough context to span multiple domains, the principal cannot oversee many heavyweight agents, and running every agent at maximum capability is economically unsustainable.

Context

A builder assembles a personal multi-agent system and faces an immediate design question: should all agents be peers, or should some agents coordinate others? The naive starting point is a flat network where any agent can call any other, or a single powerful agent that handles everything. Neither holds up as the system grows.

Forces

Solution

Separate agents into two tiers: a small number of heavyweight orchestrators with domain ownership, persistent state, and coordination authority; and a larger pool of lightweight workers that execute bounded commissions without accumulating identity across invocations. Orchestrators run on expensive, capable models; workers run on cheaper models matched to their task types. Each orchestrator owns a domain — it routes work to workers rather than executing everything directly.

Consequences

The orchestrator-worker ratio converges empirically across independent systems: Lawson reduced from approximately 30 to 8 orchestrators; the author’s personal system operates with 3 orchestrators and 7+ workers. Both landed somewhere between a 2:1 and 5:1 workers-to-orchestrators ratio.

Adding a new domain requires adding an orchestrator, not just workers. The pattern enables model tiering: workers execute bounded tasks that match cheaper models; only orchestrators require the full judgment capability of expensive models.

A flat network or single monolithic agent exhibits predictable failure modes: context overload, oversight burden, or cost unsustainability. The two-tier structure solves all three ceilings simultaneously, though it introduces coordination overhead and requires clear domain boundaries between orchestrators.

Known Results

Lawson (2026) built OpenClaw with 8 orchestrators and approximately 35 personas, converging down from an earlier attempt at 30 agents. The author’s personal system runs 3 orchestrators with 7+ workers. Osmani and Brady’s Squad system both describe a “conductor to orchestrator” and “coordinator to specialist” progression — arriving at the same structural split from different starting contexts.

Sources

Relations