- Is switching to a multi-agent system really worth it for your company?
- What did Anthropic announce, and what really works?
- How it works technically: what does each agent do, and who controls what?
- How much do multi-agent workflows really cost?
- Three concrete scenarios: what changes for an e-commerce, a professional firm, and a manufacturing company?
- What tools exist today for multi-agent orchestration, and how do they compare?
- What rules and risks do you need to know before launching a multi-agent workflow?
- Why everyone is talking about it now: is the market really ready for multi-agent orchestration?
- How to get started in practice: the steps to launch a multi-agent workflow
- What are the most common mistakes when starting with multi-agent workflows?
- How can you tell if the multi-agent workflow is working for your company?
- So, is adopting multi-agent workflows worth it for your company?
- How we see it at SHM
- Frequently Asked Questions
- Claude can now coordinate up to 1,000 AI agents in parallel with the new dynamic workflows: a main agent plans, assigns tasks and collects the results.
- Multi-agent workflows find twice as many errors as a single agent in Anthropic's tests, but they consume 5 to 30 times more tokens: calculate the cost before you start.
- Over 57% of organizations already use AI agents for multi-step workflows, but the pilot success rate remains below 15%: governance and integration are the deciding factors.
- Starting August 2, 2026, the EU AI Act imposes specific obligations on high-risk AI systems, including multi-agent workflows that touch significant decisions.
A single AI agent finds at most 27 bugs in 116,000 lines of code. A multi-agent system finds 66, more than double, in the same session. If that holds for software quality control, it holds for any repetitive, highly variable business process.
Is switching to a multi-agent system really worth it for your company?
It depends on the type of process: multi-agent workflows bring concrete benefits on parallel, highly variable tasks, but the token cost can be significant. Before deciding, you need to understand whether the quality gain justifies the expense.
A multi-agent workflow, meaning a system where a main agent coordinates dozens or hundreds of specialized agents working in parallel, isn't the right solution for every company. It works well on processes that can be broken into independent sub-tasks: quality control, analysis of large volumes of data, document review, automated testing.
On linear, simple processes, instead, it adds cost without adding value. The table below helps you figure out which situation you're in.
| Your situation | Is it worth it? | why |
|---|---|---|
| Repetitive processes with many parallel sub-tasks (e.g. QA, document audits) | Yes | The system distributes the work and cuts turnaround time; quality improves noticeably |
| Analysis of large volumes of data from multiple angles | Yes | Specialized agents work in parallel on different segments |
| Linear, simple processes (e.g. replying to an email) | No, for now | A single agent or a chatbot is cheaper and just as effective |
| Workflows with sensitive data or high-risk decisions | Proceed with caution | You need audit trails and human oversight to comply with the EU AI Act |
| First pilot with no experience with AI agents | Start small | The pilot success rate is below 15%: it's better to start from a limited use case |
What did Anthropic announce, and what really works?
Anthropic has added dynamic workflows to the Claude Managed Agents platform: a main agent creates a plan, assigns tasks to up to 1,000 agents in parallel and collects the results. In internal tests, this approach found 66 bugs in 116,000 lines of code versus the 14-27 found by a single agent.
The new feature is called dynamic workflow: instead of a single agent working in sequence, a main agent, known as the lead agent, plans the work, distributes it to dozens or hundreds of secondary agents, then collects and merges the results. All in parallel, all in the same session.
Anthropic made this feature available in public beta on May 6, 2026, along with two complementary tools: Outcomes (a separate agent that evaluates output quality against criteria defined by the user) and Dreaming (an experimental feature to reduce recurring errors across sessions). Technical details are in the developer documentation.
The most cited test: the Anthropic team hid 70 bugs in a 116,000-line codebase. A single agent found between 14 and 27 per session. The multi-agent workflow consistently found 66. A clear leap in quality, even though results on other types of tasks remain to be verified case by case, as the source itself admits (The Decoder).
| Feature | What it does | Availability |
|---|---|---|
| Dynamic Workflow | Main agent coordinates up to 1,000 agents in parallel | Public beta (since May 2026) |
| Outcomes | Separate agent evaluates output quality against defined criteria | Public beta |
| Dreaming | Reduces recurring errors across successive sessions | Research preview |
| Multi-agent sessions | Each agent has its own isolated thread; max 25 concurrent threads per session | Available |
How it works technically: what does each agent do, and who controls what?
Each agent runs in an isolated thread with its own history, but all share the same work environment. The practical limit is 25 threads active at the same time per session, not 1,000 all together.
One important detail that often gets lost in the headlines: the 1,000-agent limit applies to the workflow's total execution, not to the number of agents active at the same time. According to Anthropic's official documentation, the maximum number of concurrent threads per session is 25, main thread included.
Each agent keeps its own separate conversation history, but shares the same working environment with the others: filesystem, credentials and tools. This simplifies setup, but calls for care when you work with sensitive data.
The platform supports three coordination modes, described in the orchestration documentation:
- 1 · Lead agentPlanningThe main agent analyzes the task and builds a plan: it splits the work into independent sub-tasks and decides which agents to activate.
- 2 · Specialized agentsParallelizationIndependent sub-tasks are distributed to multiple agents working at the same time, cutting total turnaround.
- 3 · Agents with dedicated system promptsSpecializationSome agents receive specific instructions (e.g. a security agent, a documentation agent) and handle only their own domain.
- 4 · More capable agentEscalationIf a sub-task is too complex, it is handed off to a more powerful model. This raises quality, but also cost.
- 5 · Lead agentCollection and synthesisThe main agent gathers all the results, merges them and produces the final output. Outcomes can assess its quality automatically.
How much do multi-agent workflows really cost?
Multi-agent workflows consume 5 to 30 times more tokens than a single chatbot call. With premium models the cost can become significant: you need an accurate calculation before going into production.
The token is the unit AI models use to process text: every word, every symbol, every instruction has a cost. In a multi-agent workflow, each agent plans, retrieves context, calls tools, validates its output and retries failed steps. All of this multiplies the number of tokens used compared to a simple question to a chatbot.
According to CockroachLabs, the multiplier ranges from 5x to 30x compared to a single standard call. Model choice makes a huge difference: prices vary by tens of times between an economy model and a premium one, as the table below shows.
| Model | Output cost ($/M tokens) | Recommended use |
|---|---|---|
| GPT-5.6 Luna (budget) | $1,20 | Simple tasks, high volume, low criticality |
| Claude Haiku 4.5 | $5,00 | Secondary agents on standardized tasks |
| Claude Sonnet 5 | $10,00 | Quality/cost balance for most workflows |
| Claude Opus 5 | $25,00 | Lead agent or escalation on highly complex tasks |
| OpenAI o3-pro | $80,00 | Advanced reasoning; very high cost for long workflows |
Illustrative example (the numbers are meant to show the calculation, not real client data): an e-commerce store that uses Claude Sonnet 5 to review 1,000 product pages with a multi-agent workflow could consume 10-50 million output tokens per session. At $10/M tokens, that means $100-500 per session. If the workflow replaces 8 hours of manual work, the math works. If the process is already fast, it probably doesn't.
Three concrete scenarios: what changes for an e-commerce, a professional firm, and a manufacturing company?
Multi-agent workflows bring different benefits depending on the industry and the type of process. Here are three typical cases with an estimate of the value generated.
| Company type | Process that can be automated | Expected advantage | Main risk |
|---|---|---|---|
| E-commerce (200+ products) | Parallel review and optimization of product pages | Turnaround cut from days to hours; consistent tone | High token cost on large catalogs; human oversight needed |
| Professional firm (legal/tax) | Document analysis across large volumes of contracts or case files | Broader coverage, fewer errors of omission | Sensitive data: audit trail and EU AI Act compliance required |
| Manufacturing company | Automated QA on technical specifications and production reports | Anomaly detection across multiple lines in parallel | Integration with existing systems; initial setup cost |
To dig deeper into how companies are already applying these systems in real-world settings, the article on AI agents in business workflows with three real case studies offers concrete examples of adoption.
Which tools exist today for multi-agent orchestration, and how do they compare?
Claude isn't the only system that lets you coordinate multiple AI agents in parallel. There are alternatives with different features and costs: the right choice depends on the ecosystem you already use and the level of customization you need.
| platform | Max parallel agents | Strength | Main limitation |
|---|---|---|---|
| Claude Managed Agents (Anthropic) | Up to 1,000 (25 active threads) | Native integration, Outcomes for automated QA | High token cost with premium models |
| OpenAI Agents SDK | Variable (depends on the implementation) | Broad ecosystem, multiple models | A senior OpenAI engineer called 'agent swarms' a waste of tokens |
| LangGraph / LangChain | Unlimited (self-hosted) | Maximum flexibility and control | Requires custom development and advanced technical skills |
| Microsoft AutoGen | Variable | Integration with Azure and Microsoft tools | Steep learning curve |
Choosing a platform isn't just a technical decision. If your company already uses Microsoft tools, AutoGen may integrate more easily. If you want to get started fast without custom development, Claude Managed Agents offers a guided path. If you want a better picture of the AI agent landscape, the article on AI Agents and business workflows: what changes in 2026 is a good place to start.
What rules and risks do you need to know before launching a multi-agent workflow?
As of August 2, 2026, the main obligations of the EU AI Act for high-risk AI systems are in force. Multi-agent workflows that touch significant decisions — hiring, credit, safety — fall into this category and require specific measures.
The EU AI Act (the European regulation on artificial intelligence) requires high-risk systems to meet three main obligations that took effect on August 2, 2026: a record of every decision made by the system (audit trail), an immediate shutdown mechanism that a human can operate (the so-called 'stop button') and an ongoing risk management process. Iusletter explains it in detail.
In a multi-agent architecture, every agent performing a high-risk function must meet these obligations independently. It's not enough for the system as a whole to be compliant: every node in the chain must be.
| EU AI Act obligation | What this means in practice | Who is involved |
|---|---|---|
| Audit trail (Art. 12) | Record every decision made by every agent, with timestamp and rationale | IT, Legal, DPO |
| Human oversight with stop button (Art. 14) | A human operator must be able to halt the workflow at any time | Operations, IT |
| Continuous risk management (Art. 9) | Assess and update the system's risks over time, not just at launch | Management, Legal |
| Compliance for every agent (Recital 99-100) | In multi-agent architectures, every high-risk agent must be compliant | IT, Legal |
If your workflow covers low-risk internal processes, such as proofreading texts or analyzing aggregated, non-personal data, the obligations are less strict. Still, it's always smart to document your choices and keep a human in charge of the results.
Why it's in the spotlight now: is the market really ready for multi-agent orchestration?
More than half of organizations already use AI agents for multi-step workflows, but the success rate of pilots remains low. The market is growing fast, but operational maturity is still limited.
The most important figure isn't adoption, it's success: fewer than one in seven pilots reaches production, according to Fifthrow. What makes the difference is clear governance, integration with existing systems and regulatory compliance, not the choice of AI model.
To see how other tools position themselves in this market, our article on ChatGPT Work Agent and multi-app automation offers a useful comparison from an operational marketing perspective.
How to get started in practice: the steps to launch a multi-agent workflow
Anthropic recommends starting small, with a limited, well-defined use case. Here's the recommended path for an SME starting from scratch.
- 1 · Management / OperationsChoose a limited use caseIdentify a repetitive process with parallel subtasks and verifiable output. Avoid high-risk processes for your first pilot.
- 2 · IT / DevelopersEnable the right agent typeSelect 'multiagent_20261001' as the agent type on the Claude platform. Follow the official documentation or run '/claude-api managed-agents-onboard' in Claude Code.
- 3 · IT / OperationsSet token and cost limitsConfigure a maximum budget per session before launching the workflow. Multi-agent workflows can burn through tokens quickly.
- 4 · Internal teamValidate the output manually in the first sessionsDon't rely on full automation right away. Use Outcomes for automatic evaluation, but keep a human in charge of critical results.
- 5 · Management / LegalCheck regulatory complianceIf the process touches personal data or significant decisions, check your EU AI Act obligations before going into production.
- 6 · ManagementMeasure and scaleAfter the pilot, compare quality, time and costs with the previous process. Only if the balance is positive, extend it to other processes.
What are the most common mistakes when starting with multi-agent workflows?
Most pilots fail not because of technical problems, but because of poor organizational choices: overly complex processes, uncontrolled budgets and no human oversight.
Anyone approaching multi-agent orchestration for the first time tends to make the same mistakes. Knowing them in advance reduces the risk of wasting time and budget.
| Common mistake | Why it happens | How to avoid it |
|---|---|---|
| Starting with a process that's too complex | The urge to show maximum potential right away | Choose a use case with verifiable output and low risk |
| Not setting token spending limits | The consumption of agentic workflows is underestimated (up to 30x) | Configure a maximum budget per session before you start |
| Removing human oversight immediately | Aiming for total automation from day one | Keep a human in charge of the results for at least the first few weeks |
| Ignoring regulatory compliance | The assumption that it only concerns large companies | Check your EU AI Act obligations even for internal pilots |
| Always choosing the most powerful (and expensive) model | Confusing quality with model cost | Use inexpensive models for secondary agents, premium ones only for the lead agent or escalation |
| Not measuring pilot results | Carrying on out of inertia, without data | Define success metrics before you start and compare them afterward |
For a useful deep dive on how to structure an agentic workflow properly before launching it, see our article on Claude Opus 4.8 and dynamic workflows for SMEs, which explores the basics of this architecture.
How can you tell if the multi-agent workflow is working for your company?
You need concrete indicators, measured before and after the pilot. Output quality, cost per unit processed and time saved are the three key parameters.
| Indicator | How it is measured | Positive signal |
|---|---|---|
| output quality | Errors found or tasks completed correctly vs. the previous process | Improvement of ≥20% over the previous method |
| Cost per unit processed | Total token cost divided by the number of outputs produced | Lower than the cost of equivalent manual work |
| Completion time | Hours or minutes to complete the process end to end | Significant reduction compared with the manual or sequential process |
| Residual error rate | Outputs that need manual correction after the workflow | Under 10% of total outputs |
| Regulatory compliance | Complete audit trail, working stop button, up-to-date risk log | No documentation gaps in periodic checks |
So, is adopting multi-agent workflows worth it for your company?
It's worth it if you have repetitive processes with many parallel subtasks (QA, document analysis, review of large volumes), a technical team able to set up and monitor the system, and you're willing to invest in a pilot with clear metrics.
Somewhere in between if your process is semi-structured and you'd like to automate it, but you don't have experience with AI agents yet. Start with a simple single-agent workflow, measure the results, then decide whether to add multi-agent coordination.
Not yet if your process is linear and simple, your volumes are low, or you don't have the resources to handle regulatory compliance and ongoing monitoring. A chatbot or a single agent is cheaper and just as effective.
How we see it at SHM
To us, multi-agent orchestration is a powerful tool, but only when it's used to free people from routine, not to chase technological novelty. What matters isn't how many agents you can run in parallel, but how many hours of repetitive work you can give back to your team for activities that truly call for human judgment.
We've worked with manufacturing and consulting companies that handled large volumes of technical documentation: in those contexts, parallelizing agents makes sense because the process is already structured and the output is verifiable. In other cases, like creative processes, customer relationships and strategic decisions, a single agent or partial automation is the smarter choice.
If you're evaluating where and how to introduce agentic automation in your company, our strategic AI integration service always starts from an analysis of your real processes, not from a list of tools. If you want to train your internal team, our business process automation training program is designed for people without a technical background. You can also explore the topic of AI and process automation in our dedicated cluster, where you'll find up-to-date guides and analysis.
Frequently Asked Questions
The stated limit is 1,000 agents per overall workflow run, but the number of threads active at the same time per session is 25, main thread included. In practice, this means the system works in successive waves, not everything in parallel at the same instant.
Multi-agent workflows consume 5 to 30 times more tokens than a single chatbot call, according to CockroachLabs. The cost also depends on the model you choose: Claude Sonnet 5 costs $10 per million output tokens, while Claude Opus 5 goes up to $25. Before launching a workflow in production, it's essential to estimate the cost per unit of work and compare it with the value generated.
It depends on the type of process. If the workflow touches high-risk decisions, such as hiring, access to credit or safety, it falls under the obligations of Title III of the EU AI Act, in force since August 2, 2026. In multi-agent architectures, every agent that performs a high-risk function must be independently compliant.
According to Fifthrow, fewer than 15% of enterprise agentic AI pilots make it into production. What makes the difference isn't technical, it's organizational: clear governance, integration with existing systems and regulatory compliance. Starting with a narrow, well-defined use case significantly improves the odds of success.
To activate Claude's dynamic workflows, you need to access the Claude Managed Agents platform and set up the right type of agent. You don't need an advanced development team for the simpler use cases, but it's a good idea to have at least one person with basic technical skills to handle setup, cost monitoring and error management.
A process is a good fit if it can be split into independent subtasks that can run in parallel, if the output can be objectively verified and if the volumes justify the cost of tokens. Linear, simple or low-volume processes are better handled by a single agent or a traditional chatbot.
The real question isn't how many agents you can run, but how many hours of repetitive work you can give back to your people.
Related articles
Discover more articles exploring similar topics, selected to offer you a more complete and stimulating perspective. Each piece of content is carefully chosen to enrich your experience.
