10/13/25, 2:38 PM Single vs Multi-Agent System?
PHILSCHMID SEARCH ⌘K BLOG PROJECTS NEWSLETTER ABOUT ME
Single vs Multi-Agent System?
June 20, 2025 4 minute read
The debate on whats the right way to build Agents in the AI community heated up
with Cognition's "Don't Build Multi-Agents" and Anthropic's "How we built our
multi-agent research system". Despite their opposing titles both are surprisingly
aligned. The choice between a single agent and a multi-agent system isn't about
ideology, it is about picking the right tool for the right job.
Before we take a closer look, let's establish a simple definition.
“AI agent is a system that uses a Large Language Model (LLM) as reasoning
engine to decide the control flow of an application.”
What Are Single Agents?
A single-agent system operates as a “single process”. Think of it as one highly-
focused worker tackling a task from start to finish. It maintains a continuous thread
[Link] 1/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
of thought (memory) and action (tools) to ensure that every step is informed by
everything that came before it.
[Credits: Cognition Don’t Build Multi-Agents]
[Link] 2/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
Key characteristics:
Sequential: Actions are performed one after another. The agent completes step
A before moving to step B.
Unified Contenxt: Maintains a single, continuous history of the entire
conversation. Every new step has access to all previous steps, thoughts, and
tool outputs.
Stateful: Decisions made early on directly and fully inform later actions without
the need for message passing.
Pros:
Context Continuity: No information is lost between steps
Simplicity: Easier to debug, test, and maintain
Transparency: Clear execution path and decision trail
Cons:
Sequential Bottlenecks: Slow for tasks where parts could be handled in parallel.
Context Window Limitations: Eventually exceed the (useful) context window,
leading to errors and forgotten details.
Inefficiency: May waste tokens on repetitive context and are limited to single
model capabilities/instructions
[Link] 3/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
What Are Multi-Agent Systems?
A multi-agent system is structured like a team. It typically involves an “lead agent"
that breaks down a goal into smaller subtasks, which it then delegates to multiple
"worker" agents that can operate in parallel.
Key characteristics:
Parallel Execution: Subtasks can be handled simultaneously by multiple agents.
Delegation: A "lead" agent typically decomposes the main goal, delegates
subtasks, and synthesizes the results from the worker agents.
Distributed Context: Each agent operates with its own context, which is often a
subset of the total information.
Pros:
[Link] 4/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
Parallelization: Can explore multiple paths simultaneously, reducing latency.
Specialization: Each agent can be optimized and instructed for specific tasks
Breadth: Can solve complex, multi-faceted problems.
Cons:
Context: Sharing the right context between agents is hard
Coordination: Agents may duplicate work or make conflicting decisions
Cost: Can be more token-intensive. (Anthropic: 15x more tokens compared to a
standard chat interaction)
Note: It's worth mentioning other multi-agent pattern, e.g. swarm are not using a
"lead" agent. Instead they collaborate in a more peer-to-peer fashion to solve a
problem, which has both single agent and multi agent characteristics, e.g. Unified
Context, but individual instructions but comes with its on challenges.
Comparison Single Agents vs Multi-Agent Systems
Aspect Single Agent System Multi-Agent System
Context Continuous, no loss Complex sharing required
Management
[Link] 5/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
Aspect Single Agent System Multi-Agent System
Execution Sequential Parallel
Speed
Token usage 4x chat tokens 15x chat tokens
Reliability High, predictable Lower, emergent behaviors
Debugging Straightforward Complex, non-deterministic
Best For Sequential, state-dependent Parallelizable, exploratory tasks ("read"
tasks ("write" tasks). tasks).
Coordination None needed Critical success factor
Example Use Refactoring a codebase, Researching a broad market trend,
Case writing a detailed document. identifying all board members of the S&P
500.
Core Strength Context Continuity & Reliability Parallelism & Scalability
Primary Context Window Limits & Context Fragmentation & Coordination
Challenge Sequential Speed Complexity
The Universal Truths of Building Agents
Despite their opposing common principles apply for building serious Agentic
system.
[Link] 6/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
1. Context Engineering is everything: Architecting systems that dynamically
maintain the right information at the right time for reliable decision-making is the
key to success and reliability. This isn't just prompt engineering.
2. "Read" vs "Write" Agents: Are important distinction isn't single vs multi-agent.
it is whether your task primarily involves reading (research, analysis,
information gathering) or writing (code generation, content creation, file
editing).
Read tasks are easier to parallelize and suite multi-agent better
Write tasks create coordination problems when parallelized, favoring single
agents
Mixed tasks should separate read and write phases architecturally
3. Reliable Agents Need Different Tooling: Building reliable agents requires more
than a good model. It requires a robust infrastructure for durable execution to
survive failures, observability to debug behavior, and evaluation to measure
what actually matters.
4. Economic Viability and Model Improvements: Models themselves are improving
at an incredible rate. Don't over-engineer a solution for today that a much
simpler can solve for you tomorrow.
Thanks for reading! If you have any questions or feedback, please let me know on
Twitter or LinkedIn.
[Link] 7/8
10/13/25, 2:38 PM Single vs Multi-Agent System?
PHILIPP SCHMID © 2025 IMPRINT RSS FEED THEME
[Link] 8/8
