- 01. Why every handoff between agents loses approximately 37% of context -- and the three specific types of information that degrade at each transfer.
- 02. The economics of multi-agent systems: 15x token consumption compared to single-agent, and the cost curves that kill projects at scale.
- 03. The production patterns that actually work for multi-agent coordination -- and why the simplest valid architecture almost always wins.
Every Handoff Loses 37% of Context
In late 2024, Flipkart -- India's largest e-commerce platform, serving hundreds of millions of users -- deployed a multi-agent customer service system. The architecture looked clean on the whiteboard: an orchestrator agent routing to specialists for order tracking, refund processing, product recommendations, and escalation management. Each specialist had its own tool set, context, and fine-tuned personality.
Day one cost was supposed to be $3,600. Actual cost: $18,000. A 5x overrun on the first day of production.
The post-mortem uncovered four compounding problems. Full conversation history shipped at every handoff -- 15,000 tokens per transfer, mostly irrelevant. The orchestrator made four LLM calls per routing decision. No caching existed. And 18% of interactions entered infinite retry loops where the orchestrator couldn't determine which specialist to invoke.
The team rewrote the system in three weeks. The new architecture: a single agent with dynamic tool loading. One context window. No handoffs. No infinite routing loops. Cost dropped to below original budget. Customer satisfaction scores were higher.
Coordination Patterns and Their Loss Curves
Multi-agent coordination is the practice of splitting a task across multiple AI agents, each operating with its own context, tools, and reasoning, connected through some form of handoff protocol. The appeal is intuitive: just as human organizations divide work across specialists, AI systems should divide work across specialist agents.
The reality is brutal. Research across 1,600+ execution traces spanning seven major multi-agent frameworks revealed failure rates between 40% and 80%. The Multi-Agent System Failure Taxonomy categorized 14 distinct failure modes into three buckets: system design issues (42%), inter-agent misalignment (37%), and task verification gaps (21%).
At every handoff between agents, three types of information degrade. State loss: the receiving agent gets a summary, not the reasoning chain. Intent loss: the customer's original meaning gets diluted to neutral labels. Reasoning loss: failed approaches aren't transferred, so downstream agents repeat them.
Multi-agent coordination is a distributed systems problem wearing an AI costume. Every problem that distributed computing solved over 40 years reappears here.
-- The structural insight~37% of Context.
When multi-agent actually earns its cost. Three conditions justify splitting into multiple agents. Capability separation: the agents genuinely need different tool sets that would create security conflicts in a single agent. Evaluator independence: the agent that generates output should be separate from the agent that evaluates it. Scale-driven parallelism: genuinely independent sub-tasks that benefit from parallel execution.
The patterns that survive production. Two coordination architectures have proven reliable at enterprise scale. Orchestrator-worker with structured handoff packets -- packets that include original intent verbatim, work completed with reasoning, key facts curated, open questions explicit, and confidence flags. And shared state stores where all agents read from and write to a common state object, eliminating state loss but introducing concurrency challenges.
When Multi-Agent Beats Single-Agent (And When It Doesn't)
The multi-agent demo trap. Multi-agent architectures produce impressive demonstrations. A research agent finds information, a writing agent drafts, a review agent critiques. In production, research findings lose nuance at the handoff. The writing agent produces content missing key caveats. The demo sells the architecture. Production reveals the handoff tax.
Compliance and auditability. When three agents contribute to a decision, the reasoning is distributed across three separate traces. Reconstructing "why did the system recommend X?" requires stitching together three separate execution logs. Some regulatory frameworks require end-to-end explanation of any automated decision.
The Trap
"More agents = better."
The intuition runs deep: just as a team of human specialists outperforms a single generalist, a team of AI specialists should outperform a single general agent. This analogy is seductive and wrong in one critical way: human specialists share years of collaborative context. AI agents share none of this. Every handoff is a cold introduction between strangers.
The fix: before building multi-agent, prove -- with data, not intuition -- that a single agent fails at the task.The Handoff Contract
Consider a large property insurance company processing 8,000 claims per day. The original multi-agent architecture used five agents connected by four handoffs. Token cost per claim: 85,000 tokens. Daily cost: $3,360. And 23% of claims required human re-review because context degraded across the pipeline.
Lost More Than It Earned.
Remember This
1. Every agent handoff loses approximately 37% of context through three mechanisms: state loss, intent loss, and reasoning loss. This isn't a bug to fix -- it's a structural property of transferring information between separate context windows.
2. Multi-agent systems consume 15x the tokens of single-agent systems. At 100,000 queries per day, the difference is $140K daily. The architecture must produce measurably better outcomes to justify this cost.
3. Three conditions justify multi-agent: capability separation for safety, evaluator independence, and genuine parallelism. If none apply, use a single agent.
References
1. Building Effective Agents -- Anthropic Engineering Blog
2. Are Multi-Agent Systems the Future? -- MAST Research
3. OpenAI Agents SDK: Handoffs -- OpenAI Documentation