Single Agent vs Multi-Agent AI for Australian Businesses

A black chess king stands in focus beside a black bishop and a white pawn on a glossy chessboard, with other pieces blurred behind.

Single agent vs multi-agent is now one of the first design questions anyone building an AI agent has to answer: do you give one agent all the tools, or split the work across a team of specialist agents run by a coordinator? Two announcements in the past month pull in different directions. Anthropic has published guidance on folding most subagents back into one main agent, and Cognition has launched Devin Bots, agents that lead teams of other agents.

Here’s what an AI agent is, what each company actually said, and how an Australian business can pick the right design without over-building.

What is an AI agent and how does it work?

An AI agent is a language model that can decide its own next step. You give it a goal, some instructions and a set of tools (search your documents, look up an order, draft an email), and it works in a loop: it picks a tool, reads the result, then decides whether to use another tool, ask you a question or give an answer.

That is the main way agentic AI differs from traditional automation. Anthropic’s earlier guide, Building effective agents, draws the line clearly: workflows are “orchestrated through predefined code paths”, while agents “dynamically direct their own processes and tool usage”. If every step is the same every time, a fixed automation is usually simpler and cheaper to run. If a step needs judgement, an agent may fit. Our guide to automation vs agents covers that choice in more depth, and the free AI agents course walks through how agents work from the ground up.

Single agent vs multi-agent: what’s the difference?

  • Single agent. One model, one conversation and one context window. It holds all the tools and instructions it needs, and loads detailed instructions only when a task calls for them.
  • Multi-agent. A coordinator (often called an orchestrator) hands pieces of work to subagents. Each subagent has its own instructions, tools and context window, and sends a result back to the coordinator.

Multi-agent designs can look tidier on a whiteboard, because each team or task gets its own agent. The trade-off is that every handoff passes a summary rather than the full picture, and every extra agent adds another model call.

What Anthropic’s guide says about single agent vs multi-agent

Anthropic’s article Building agents: When to fold subagents into your main agent loop (dated 22 October 2026) is written for teams building on its Claude Platform, and it has been widely shared this week. The key points:

  • One agent with skills often did better. Across several enterprise deployments it compared, Anthropic says a single agent with skills “frequently outperformed” both a one-big-prompt design and a subagent design on quality, and “often at a lower cost and latency per task”.
  • The speed claim is one customer’s result. Anthropic says one customer that compared its multi-agent setup with a single agent loop saw “a roughly 2x faster runtime” with the single loop. That is one customer’s measurement, not a general rule that single agents are twice as fast.
  • Agents shouldn’t copy your org chart. The guide warns against giving every team or problem type its own subagent, which it says leads to lost detail in handoffs, overlapping agents and conflicting instructions.
  • Folding in can break things. In Anthropic’s own test on a sample support agent, folding one subagent into another lowered latency but broke a process that had been passing. The break only showed up in one of two runs.

Anthropic is also clear that subagents still have a place. It suggests keeping them when a single agent measurably degrades because its context gets too long, when several tasks need to run in parallel with fresh context, or when one agent struggles to pick the right tool or has conflicting instructions.

What Cognition’s Devin Bots are

On 17 November (US time), Cognition introduced Devin Bots, which it describes as “a persistent agent with its own memory and tools that can lead multiple Devins in parallel”. Devin is Cognition’s AI software engineer, so Devin Bots are built for software teams, not general office work.

According to the Devin Bots documentation, you give a Bot an outcome to own. It plans the work, starts a separate Devin session for each piece, follows up on those sessions and collects anything that needs a person (decisions, pull requests to review, requests for new access) into one “Needs you” list. A Bot asks before it takes on new access, and its usage counts toward your plan. Cognition’s pricing page lists Devin Bots on its Max, Teams and Enterprise plans.

Do these two announcements contradict each other?

Not really. Anthropic’s advice is mostly about agents that hold one conversation with a customer or staff member, where splitting the work across many specialists can hurt quality and consistency. Devin Bots are aimed at long-running engineering work that can be broken into independent pieces, which is the parallel-work case where Anthropic says subagents still make sense.

The shared lesson is to split work by what each piece needs to know, not by who owns it. Anthropic’s guide recommends giving each subagent “a piece of work that needs little from the rest of the system” and having it hand back a compact result.

Which design suits an Australian business?

For most small and mid-sized businesses starting out, a single agent is the sensible default. A practical way to decide:

  • Start with one agent if the job is one conversation or one process, such as answering customer questions from your policies, triaging an inbox or preparing a quote for review.
  • Consider more than one agent only when you can show a single agent is struggling, when separate tasks genuinely run at the same time and don’t depend on each other, or when parts of the work need different access and permissions.
  • Use plain automation for the fixed steps around the agent, such as copying a form into your CRM. See our AI automation service.
  • Consider whether you need an agent at all. If staff mainly need accurate answers from your documents, a knowledge assistant may be enough. Our guides on AI agent vs chatbot and RAG vs AI agents explain the difference.

Much of what Anthropic suggests moving into a single agent is written instructions: skills that describe how your team does a task. Writing those well is a prompting skill, which our free AI prompts course covers.

Keep people in control, whichever design you choose

More agents means more places for an action to happen without anyone noticing. The Australian Government’s Guidance for AI Adoption from the National AI Centre sets out six essential practices, starting with deciding who is accountable and including testing and monitoring and maintaining human control. In practice, for any agent design:

  • Have a person approve anything that sends, pays, deletes or changes customer records.
  • Give each agent only the access it needs, and keep a log of what it did.
  • Name one person who owns the agent’s overall behaviour, rather than each team owning a piece.

We covered approvals and audit trails for agents in more detail in our post on AI agent safety in Australia.

How to test single agent vs multi-agent before you commit

Anthropic’s main advice is to measure rather than guess. Its guide suggests:

  1. Build a small set of real test cases first. It says “even a set of 10 golden examples is enough to start”.
  2. Change one thing at a time, for example folding in one subagent, and compare results.
  3. Look at which tasks changed from pass to fail, not just the overall score.
  4. Run each test more than once, because some failures only show up occasionally.

The same approach works for a small business deciding whether to add a second agent: write down ten typical jobs, check the current setup against them, and only add complexity if the results improve.

Caveats

  • Vendor advice. Anthropic and Cognition both sell agent products. Their guidance reflects their own platforms and customers.
  • No single answer. Neither source says one design always wins. Anthropic says the right choice depends on the task and should be tested.
  • Devin Bots are new. They launched this week for software engineering teams, and plan details may change.
  • No guaranteed savings. Whether an agent saves time or money depends on your process, your data and how carefully it is tested and supervised.

This is general information based on Anthropic’s published guidance, Cognition’s announcement, documentation and pricing page, and the National AI Centre’s Guidance for AI Adoption, as at 18 November 2026. Products, plans and recommendations in this area change often. It is not legal or technical advice for your specific situation.

Planning an AI agent for your business?

Start with our free AI agents course to learn how agents work and where they fit. If you’d like a second opinion, an AI business audit shows where agents or simpler automation make sense in your operations. Aideveloper designs and builds AI agents for Australian businesses, with human approval steps, limited access and testing built in. Talk to us about whether one agent or several suits your work.

Featured image: Photo by Pixel Senses on Pexels.