Multi-Agent Orchestration: The 2026 Cost Reality

Multi-agent systems are 5–15× more expensive per user request than single-agent equivalents. They're worth it — sometimes.

The context-inheritance tax

When a supervisor agent spawns three sub-agents, each sub-agent typically inherits:

• The system prompt (2–8K tokens)
• The tool schemas (5–20K tokens)
• The conversation history so far (variable)
• The supervisor's plan / task assignment (0.5–2K tokens)

That prefix is paid once per sub-agent, per tool call. A 3-agent system doing 5 tool calls each = 15× the prefix cost of the same work done sequentially.

The parallel-vs-serial trade

Architecture	Latency	Cost	When to use
Single agent, serial	High	1× (baseline)	Simple pipelines; cost-sensitive
Single agent, parallel tools	Medium	1.2–1.5×	Independent sub-tasks; same context
Multi-agent, shared context	Medium	3–5×	Specialisation matters; context reuse
Multi-agent, isolated context	Low	8–15×	True independence; latency-critical

Anthropic's published guidance

Anthropic's 2025 multi-agent research report notes that their internal Claude research agent uses ~15× the tokens of a single-agent equivalent — and they recommend multi-agent only when:

1. The task has genuine parallelism (research, breadth-first search).
2. Sub-tasks are independent enough that context isolation is a feature, not a bug.
3. The user-facing value (latency, quality) justifies the token multiplier.

The four cost controls

1. Compress before delegation. The supervisor should hand each sub-agent a distilled task brief (200–500 tokens), not the full conversation.
2. Scope tools per sub-agent. A "search" sub-agent gets search tools only. A "write" sub-agent gets file tools only. Never expose the full toolbox to every agent.
3. Summarise before aggregation. Sub-agent outputs get summarised to 500–1000 tokens before the supervisor synthesises. Raw outputs stay in a separate store.
4. Cap fan-out. Hard limit on concurrent sub-agents (typically 3–5). Fan-out of 20 sub-agents almost always indicates a design smell.

When multi-agent is a mistake

• Sequential workflows dressed as agents. If step B needs step A's output, it's a pipeline, not a multi-agent system. Use LangGraph nodes, not autonomous agents.
• Coordination-heavy tasks. If sub-agents need to negotiate, you're paying inter-agent chatter tokens for no capability gain.
• Small-context problems. If the whole task fits in 8K tokens, one agent will do it faster and cheaper.

The observability minimum

Every span must record: agent_id, parent_agent_id, depth, tokens_input, tokens_output, tools_available_count. Without these five fields you cannot debug why a multi-agent run cost 40× what you expected — the answer is almost always "tool schemas × depth × fan-out".

Vendor direction

2026 frameworks (LangGraph 0.4+, CrewAI 2.0, Autogen 0.5) now expose hierarchical context management as a first-class concept: the parent's context is not automatically inherited; sub-agents receive an explicit context_view object the developer defines. This is the correct default. Frameworks that still inherit-by-default are becoming a liability.