---
title: "Multi-Agent Orchestration"
description: "Hierarchical agent swarms where engineers orchestrate fleets of specialized agents in parallel isolated worktrees: coder to conductor to orchestrator."
section: "concepts"
type: "concept"
created: "2026-08-24"
updated: "2026-08-30"
confidence: "high"
tags: ["agents", "workflow", "subagents", "context-engineering", "evaluation"]
canonical: "https://pyweb.dev/wiki/multi-agent-orchestration"
---
# Multi-Agent Orchestration

**Multi-Agent Orchestration** is the paradigm shift from single-agent coding to **hierarchical agent swarms** where human engineers orchestrate fleets of specialized agents working in parallel across isolated worktrees. This represents the evolution from coder → conductor → orchestrator.

## Hierarchical Architecture

```mermaid
flowchart TD
    HO["HUMAN ORCHESTRATOR<br/><small>(High-level goals & spec)</small>"] --> LP["LEAD PLANNER / AGENT<br/><small>(Frontier planner model)</small>"]
    
    LP --> SA1["SUBAGENT 1 (Work)<br/><small>Isolated worker · Worktree: feature-a</small>"]
    LP --> SA2["SUBAGENT 2 (Work)<br/><small>Isolated worker · Worktree: feature-b</small>"]
    LP --> SA3["SUBAGENT 3 (Test)<br/><small>Test/review worker · Conformance tests</small>"]
    
    SA1 --> JM["JUDGE / MERGE AGENT<br/><small>(Adversarial verification)</small>"]
    SA2 --> JM
    SA3 --> JM
```

## Context Compression via Subagents

The fundamental mathematical driver is the **context window constraint**. Complex tasks with huge codebases flood single context windows with intermediate tool outputs, causing reasoning degradation and [context rot](/wiki/context-rot).

**Subagents as Compression Engines:** The lead orchestrator spawns isolated [subagents and context management](/wiki/subagents-and-context-management) for specific exploratory or implementation subtasks. Each subagent operates in its own fresh context, runs tools in parallel, and returns only a compressed, high-signal summary back to the parent. [[source: agentic-engineering-trends-2026-synthesis]](/wiki/raw/articles/agentic-engineering-trends-2026-synthesis)

**Performance impact:** Anthropic reports that its multi-agent research system performs especially well on breadth-first tasks, but that result is workload-specific rather than a general guarantee. [[source: agentic-engineering-trends-2026-synthesis]](/wiki/raw/articles/agentic-engineering-trends-2026-synthesis)

## Primary Source: Anthropic's Research System (2025)

Anthropic's published account of its Research feature is the canonical orchestrator-worker case study. A lead agent plans, saves the plan to external memory, and spawns parallel subagents with explicit objectives, output formats, tool guidance, and task boundaries — vague delegation ("research the semiconductor shortage") produced duplicated and misinterpreted work. [[source: anthropic-multi-agent-research-system-2025]](/wiki/raw/articles/anthropic-multi-agent-research-system-2025)

Key measured findings:

- A multi-agent system with Claude Opus 4 lead and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on Anthropic's internal research eval.
- On BrowseComp, token usage by itself explains 80% of performance variance; tool-call count and model choice explain most of the rest — multi-agent architectures win mainly by spending more tokens across separate context windows.
- Cost reality: agents use about 4× more tokens than chat, and multi-agent systems about 15× more, so the task's value must justify the burn.
- Parallelism (3-5 subagents spawned concurrently, 3+ tools called in parallel per subagent) cut research time by up to 90% for complex queries. [[source: anthropic-multi-agent-research-system-2025]](/wiki/raw/articles/anthropic-multi-agent-research-system-2025)

Operational lessons: embed effort-scaling rules in prompts (simple fact-finding = 1 agent, 3-10 tool calls); prefer end-state evaluation over prescribed step-checking; let subagents write outputs to a filesystem and pass lightweight references to avoid a "game of telephone"; use rainbow deployments because stateful agents may be mid-run during any deploy. [[source: anthropic-multi-agent-research-system-2025]](/wiki/raw/articles/anthropic-multi-agent-research-system-2025)

## Token Economics & Cost Multipliers

Running parallel agent swarms introduces significant token multiplication:

- Interactive chat is the lowest-cost baseline.
- Autonomous loops consume more context through repeated observation and action.
- Parallel subagents multiply token use again; the additional breadth must justify that cost. [[source: agentic-engineering-trends-2026-synthesis]](/wiki/raw/articles/agentic-engineering-trends-2026-synthesis)

### Economic Reality Check
The productivity gains must justify the token cost multiplication. Multi-agent patterns work economically when:
1. Independent subtasks can run concurrently.
2. The answer benefits from separate context windows.
3. Quality improvements from adversarial verification justify the overhead.

## Landmark Case Study: FastRender Browser Engine

In January 2026, the Cursor engineering team tested long-running autonomous swarms by building a web browser engine from scratch in Rust (`fastrender`):

- **Scale:** A very large generated Rust codebase spanning many files
- **Architecture:** Hierarchical tree of planners breaking browser specifications into modular tickets  
- **Execution:** Hundreds of concurrent agents over one week across isolated workspaces
- **Verification:** Continuous web conformance test suites with terminal Judge agent
- **Result:** Successfully rendered complex real-world pages (google.com, personal blogs) directly to pixels [[source: agentic-engineering-trends-2026-synthesis]](/wiki/raw/articles/agentic-engineering-trends-2026-synthesis)

## Orchestration Tooling (2026)

1. **Melty Labs Conductor:** Multiple Claude Code agents in parallel on isolated Git worktrees
2. **Claude Squad:** Multiplexes coding agents across concurrent terminal (`tmux`) panes
3. **Pi Agent Framework:** Lightweight, headless agent engine with offline execution and specialized benchmark plugins [[source: agentic-engineering-trends-2026-synthesis]](/wiki/raw/articles/agentic-engineering-trends-2026-synthesis)

## Failure Modes

| Symptom | Root cause | Fix |
|---|---|---|
| Agents duplicate work | No task partition before dispatch | Partition by ownership boundaries first ([agents vs workflows](/wiki/agents-vs-workflows)) |
| Context explode per agent | Full history broadcast to all | Give each agent only its slice + a summary pointer |
| Merge conflicts on recombination | Concurrent writes to shared state | Serialize writes through one coordinator or partition the state |

## Rule of Thumb

Add an agent only when the context split saves more tokens than the
coordination costs.

## Related Concepts
- [agent harness engineering](/wiki/agent-harness-engineering)
- [subagents and context management](/wiki/subagents-and-context-management)
- [context engineering](/wiki/context-engineering)
- [automated eval engineering](/wiki/automated-eval-engineering)
- [conformance suites as fitness functions](/wiki/conformance-suites-as-fitness-functions)

---

## Agent Navigation

cluster: person (170 pages) | betweenness: 295.4

### References (outbound)
- [Context Rot](https://pyweb.dev/wiki/context-rot.md)
- [Subagents and Context Management](https://pyweb.dev/wiki/subagents-and-context-management.md)
- [Agents vs Workflows](https://pyweb.dev/wiki/agents-vs-workflows.md)
- [Agent Harness Engineering](https://pyweb.dev/wiki/agent-harness-engineering.md)
- [Context Engineering](https://pyweb.dev/wiki/context-engineering.md)
- [Automated Eval Engineering](https://pyweb.dev/wiki/automated-eval-engineering.md)
- [Conformance Suites as Fitness Functions](https://pyweb.dev/wiki/conformance-suites-as-fitness-functions.md)

### Referenced by (inbound)
- [Agent Containment and Blast Radius](https://pyweb.dev/wiki/agent-containment-and-blast-radius.md)
- [Agentic Software Factory](https://pyweb.dev/wiki/agentic-software-factory.md)
- [Agents vs Workflows](https://pyweb.dev/wiki/agents-vs-workflows.md)
- [Conformance Suites as Fitness Functions](https://pyweb.dev/wiki/conformance-suites-as-fitness-functions.md)
- [Deterministic Lint Gates](https://pyweb.dev/wiki/deterministic-lint-gates.md)
- [Generator-Evaluator Loop](https://pyweb.dev/wiki/generator-evaluator-loop.md)
- [Anthropic](https://pyweb.dev/wiki/anthropic.md)
- [Viv Trivedy](https://pyweb.dev/wiki/viv-trivedy.md)
- [Wilson Lin](https://pyweb.dev/wiki/wilson-lin.md)

### Evidence (verified primary sources)
- [agentic-engineering-trends-2026-synthesis](https://pyweb.dev/wiki/raw/articles/agentic-engineering-trends-2026-synthesis.md) | origin: https://pyweb.dev/wiki/raw/articles/agentic-engineering-trends-2026-synthesis | ingested: 2026-08-24 | sha256: 6c78ca873a4df069f1606e9f68538de7b428633720e48e184fad0ea31b5d2a25
- [anthropic-multi-agent-research-system-2025](https://pyweb.dev/wiki/raw/articles/anthropic-multi-agent-research-system-2025.md) | origin: https://www.anthropic.com/engineering/multi-agent-research-system | ingested: 2026-08-24 | sha256: ea59deea6bf46c19a1b61cb74cad5da0f5ad62f4c5d2c555209538ee99566626

### Machine endpoints
- Knowledge graph: https://pyweb.dev/api/graph.json
- Graph analysis: https://pyweb.dev/api/graph-analysis.json
- Context index: https://pyweb.dev/llms.txt