Divya van Mahajan

Should You Build Single Agents or Multi Agent Systems?

ai-daily-brief-podcast

Should You Build Single Agents or Multi-Agent Systems?

Overview

This episode of the AI Daily Brief (published June 18, 2025) examines a current architectural debate in AI engineering: whether to build single-agent or multi-agent systems for complex tasks. The episode synthesizes two contrasting published perspectives — one from Anthropic (“How We Built Our Multi-Agent Research System”) and one from Cognition (makers of the Devin AI coding tool, authored by Walden Yan, titled “Simply Don’t Build Multi-Agents”) — to help practitioners think through which approach suits their use case.

The episode also covers two headline stories: escalating tensions between OpenAI and Microsoft over OpenAI’s corporate restructuring, and the fallout from Meta’s acquisition of a 49% stake in Scale AI.

Source video URL: (not provided)

Prerequisites

  • Basic familiarity with large language models (LLMs) and how they process prompts and context
  • Understanding of what an AI agent is — an LLM-powered system that autonomously takes actions over multiple steps to complete a task
  • Awareness of concepts like context windows, token usage, and API-based model access
  • Familiarity with the general landscape of AI companies (Anthropic, OpenAI, Microsoft, Meta, etc.)
  • Some exposure to software engineering concepts such as parallel processing and pipeline architectures

Main Points

Headline 1: OpenAI vs. Microsoft Tensions Over Corporate Restructuring

  • OpenAI is converting from a nonprofit to a public benefit corporation; Microsoft is reportedly the only investor opposing the conversion due to concerns about its lucrative revenue- and profit-sharing agreement (20% of revenues, 49% of profits, capped at ~$120 billion).
  • OpenAI executives have reportedly discussed seeking a federal antitrust review of the Microsoft investment contract, and a PR campaign targeting Microsoft — though both companies issued a joint statement calling the partnership “productive.”
  • Key sticking points include: how many shares Microsoft receives post-conversion, OpenAI’s desire to exit its exclusive cloud-hosting agreement with Microsoft (which currently bars Amazon and Google from selling OpenAI models via API), and whether Microsoft gains access to Windsurf’s technology via the pending acquisition.
  • OpenAI has offered Microsoft a 33% stake in the restructured entity — a significant reduction from the current 49% profit share — and faces a year-end deadline to complete restructuring or risk losing $20 billion in committed funding.

Headline 2: Google and Others Cut Ties with Scale AI Following Meta Deal

  • Reuters reported that Google — Scale AI’s largest customer, accounting for ~$150 million of Scale’s $870 million in 2024 revenue — is ending its relationship with Scale following Meta’s acquisition of a 49% stake.
  • Microsoft and xAI were also reportedly planning to end contracts with Scale, raising questions about what Meta actually acquired for ~$14.3 billion.
  • The deal structure suggests Meta’s primary goal was securing Scale CEO Alexander Wang to lead Meta’s superintelligence efforts; Wang personally stands to receive over $1 billion vesting over five years contingent on his continued employment.
  • Meta negotiated protective terms: if Scale is sold within 2.5 years, Meta is fully repaid before other investors. The deal is described as highly binary — either a stroke of genius or a massive overpayment.

Main Topic: The Case For Multi-Agent Systems (Anthropic)

  • Anthropic’s multi-agent research system uses a central lead (orchestrator) agent that breaks down a user query into discrete subtasks, then dispatches multiple sub-agents to execute those subtasks in parallel using search tools, before collating results into a final report.
  • Key benefits identified:
    • Parallelization: dozens of sub-agents work simultaneously, dramatically increasing speed.
    • Fault tolerance: if one sub-agent fails, the orchestrator spawns a new one with adjusted instructions, rather than the whole workflow breaking.
    • Context window management: each sub-agent operates within its own context window, returning a compressed result to the lead agent, circumventing the limits of a single context window.
    • Cost-effective model mixing: a Claude 4 Opus orchestrator with Claude 4 Sonnet sub-agents outperformed a single Claude 4 Opus agent by 90.2% on internal evaluations.
  • Anthropic’s core explanation for why multi-agent systems work: “Multi-agent systems work mainly because they help spend enough tokens to solve the problem. Token usage by itself explains 80% of the improvement.”
  • Significant downside: cost. Agents use ~4× more tokens than chat interactions; multi-agent systems use ~15× more tokens than chats. The value of the task must justify the cost.
  • Anthropic explicitly notes that multi-agent systems are not universally applicable: tasks with heavy interdependencies between sub-agents (e.g., most coding tasks) are a poor fit.

Main Topic: The Case Against Multi-Agent Systems (Cognition / Walden Yan)

  • Cognition argues that parallel multi-agent architectures are fragile in practice because sub-agents lack the shared context needed to produce consistent, coherent outputs.
  • Illustrative example: a Flappy Bird clone task split between two sub-agents — one builds the background (misinterprets the style as Super Mario Bros.), the other builds the bird (wrong visual style and movement). The orchestrator then must reconcile two incompatible outputs.
  • Simply passing the original task as context to sub-agents is insufficient: in real multi-turn production systems, there are tool calls, intermediate decisions, and accumulated context that sub-agents cannot see and that materially affect correct interpretation.
  • Two core principles derived from their analysis:
    1. Share context and share full agent traces, not just individual messages.
    2. Actions carry implicit decisions; conflicting decisions produce bad results.
  • Their preferred alternative is a linear handoff (sequential) design: a single agent breaks down the task and then executes each subtask in sequence, carrying the full accumulated context forward at each step.
  • For tasks too large for a single context window, they propose a context compression side process — a separate LLM that distills the key decisions and context from prior steps into a condensed form passed to the next step in the chain.
  • Cognition acknowledges this is a snapshot in time and expresses optimism about multi-agent collaboration as model capabilities improve.

Synthesis: Which Approach Should You Use?

  • The two arguments are not directly contradictory — they address different use cases:
    • Anthropic’s multi-agent approach suits tasks where subtasks are truly independent and parallelizable (e.g., web research where each sub-agent looks up a different company).
    • Cognition’s single-agent sequential approach suits tasks where subtasks are interdependent and require shared context (e.g., coding, where visual style and logic must remain consistent across components).
  • The determining question: Are the subtasks dependent on one another, or can they be executed independently?
  • Both Anthropic and Cognition agree that multi-agent systems are not yet mature for tasks requiring real-time coordination and context sharing between agents — this is a constraint of current capability, not a permanent limitation.

Key Concepts

  • Multi-agent system: An architecture in which a lead (orchestrator) agent delegates subtasks to multiple specialized sub-agents that operate in parallel, then aggregates their outputs.
  • Single-agent (sequential/linear handoff) design: An architecture where one agent breaks down a task and executes subtasks sequentially, carrying full context forward at each step.
  • Orchestrator agent: The central agent in a multi-agent system responsible for decomposing tasks, dispatching sub-agents, and collating results.
  • Sub-agent: A specialized agent tasked with completing a discrete, bounded portion of a larger workflow.
  • Context window: The maximum amount of text (tokens) an LLM can process in a single interaction; a fundamental constraint in agent design.
  • Context engineering: The practice of dynamically constructing, managing, and compressing the context provided to an LLM agent in a running system — described as the primary job of engineers building agents in 2025.
  • Parallelization: Running multiple independent subtasks simultaneously across separate agents to improve speed and throughput.
  • Context compression: A technique where a secondary LLM summarizes the key decisions and history of a long agent session into a smaller context passed to the next step, enabling long tasks to proceed without exceeding context limits.
  • Token usage: The measure of how many tokens an LLM consumes in processing; directly correlated with cost and, per Anthropic, the primary driver of multi-agent performance gains.
  • Fault tolerance: The ability of a system to recover from individual component failures without the entire workflow breaking down.
  • Devin: An AI-powered coding agent developed by Cognition.
  • Scale AI: A data labeling and AI infrastructure company in which Meta acquired a ~49% stake for ~$14.3 billion in mid-2025.

Summary

The central message of the main episode is that the choice between single-agent and multi-agent architectures is use-case dependent, not ideological. Anthropic’s research shows that multi-agent systems excel when tasks can be decomposed into truly independent, parallelizable subtasks — as in web research — delivering dramatic speed and performance improvements largely because they allow more tokens to be spent on a problem. Cognition, drawing from their experience building a coding agent, argues that parallelism introduces fragility when sub-agents lack shared context, and advocates instead for a sequential, single-agent architecture in which full context travels linearly through the workflow. Both camps agree on the fundamental importance of context management, and both acknowledge that multi-agent coordination across interdependent tasks remains an unsolved problem at current capability levels — one that ongoing research is actively working to address.