- Is it really worth switching to a multi-agent system for your business?
- What did Anthropic announce and what actually works?
- How it works under the hood: what does each agent do and who checks what?
- How much does it really cost to use multi-agent workflows?
- 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 enabling a multi-agent workflow?
- Why are people talking about this 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 people make when starting with multi-agent workflows?
- How to tell if the multi-agent workflow is actually working for your business?
- So, is it worth your company adopting multi-agent workflows?
- How we see it at SHM
- Frequently Asked Questions
- Claude can now coordinate up to 1,000 AI agents in parallel with new dynamic workflows: a main agent plans, distributes tasks, and collects results.
- Multi-agent workflows find twice as many errors as a single agent in Anthropic's tests, but consume 5 to 30 times more tokens: the cost must be calculated before starting.
- Over 57% of organizations already use AI agents for multi-step workflows, but pilot success rates remain below 15%: governance and integration are the decisive factors.
- From August 2, 2026, the EU AI Act will impose precise obligations for high-risk AI systems, including multi-agent workflows that affect relevant 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 this holds true for software quality control, it also applies to any repetitive and high-variability business process.
Is it really worth switching to a multi-agent system for your business?
It depends on the type of process: multi-agent workflows bring concrete benefits for parallel and highly variable tasks, but the token cost can be significant. Before deciding, you need to understand if the gain in quality justifies the expense.
A multi-agent workflow — meaning a system where a main agent coordinates tens or hundreds of specialized agents working in parallel — is not the right solution for every company. It works well on processes that can be broken down into independent sub-tasks: quality control, analysis of large volumes of data, document review, automated testing.
On linear and simple processes, however, it adds cost without adding value. The table below helps to understand your situation.
| Your situation | Is it worth it? | why |
|---|---|---|
| Repetitive processes with many parallel sub-tasks (e.g. QA, document audit) | Yes | The system distributes the work and cuts down time; quality improves noticeably |
| Analysis of large data volumes from multiple angles | Yes | Specialized agents work in parallel on different segments |
| Linear and simple processes (e.g. replying to an email) | No, for now | A single agent or chatbot is cheaper and just as effective |
| Workflows with sensitive data or high-risk decisions | With caution | Audit trails and human supervision are needed to comply with the EU AI Act |
| Initial pilot without prior AI agent experience | Start small | Pilot success rates are below 15%: it's better to start with a limited use case |
What did Anthropic announce and what actually works?
Anthropic has added dynamic workflows to the Claude Managed Agents platform: a main agent creates a plan, distributes 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 compared to 14-27 with a single agent.
The new thing is called dynamic workflow (dynamic workflow): instead of a single agent working sequentially, a main agent — called lead agent — plans the work, distributes it to tens or hundreds of secondary agents, and 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 based on user-defined criteria) and Dreaming (an experimental feature to reduce recurring errors between sessions). The technical details are in the developer documentation .
The most cited test: Anthropic's 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 forward, although the results on other types of tasks remain to be verified case by case, as the source itself admits ( The Decoder ).
| Function | What it does | Availability |
|---|---|---|
| Dynamic Workflow | Main agent coordinates up to 1,000 agents in parallel | Public beta (since May 2026) |
| Outcomes | A separate agent evaluates output quality against defined criteria | Public beta |
| Dreaming | Reduces recurring errors across consecutive sessions | Research preview |
| Multi-agent sessions | Each agent has its own isolated thread; max 25 concurrent threads per session | Available |
How it works under the hood: what does each agent do and who checks what?
Each agent operates in an isolated thread with its own history, but all share the same workspace. The practical limit is 25 active threads simultaneously per session, not 1,000 all at once.
An important detail that often gets lost in the headlines: the limit of 1,000 agents concerns the overall execution of the workflow, not the number of agents active at the same time. According to the official Anthropic documentation , the maximum concurrent threads per session is 25, main thread included.
Each agent has its own separate conversation history, but shares the same operating environment with others: filesystem, credentials, and tools. This simplifies setup but requires caution when working with sensitive data.
The platform supports three coordination modes, described in the orchestration documentation :
- 1 · Lead agent Planning The main agent analyzes the task and makes a plan: it breaks the work down into independent sub-tasks and decides which agents to trigger.
- 2 · Specialized agents Parallelization Independent sub-tasks are distributed to multiple agents working simultaneously, reducing total time.
- 3 · Agents with dedicated system prompts Specialization Some agents receive specific instructions (e.g., security agent, documentation agent) and handle only their own domain.
- 4 · Most capable agent Escalation If a sub-task is too complex, it is delegated to a more powerful model. This increases quality but also cost.
- 5 · Lead agent Collection and synthesis The main agent gathers all the results, combines them, and produces the final output. Outcomes can automatically evaluate their quality.
How much does it really cost to use multi-agent workflows?
Multi-agent workflows consume 5 to 30 times more tokens than a single chatbot call. With premium models, the cost can become significant: precise calculation is needed before going into production.
The tokens is the unit of measurement with which AI models process text: every word, every symbol, every instruction has a cost. In a multi-agent workflow, each agent plans, retrieves context, calls tools, validates the 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. The choice of model has a huge impact: prices vary by tens of times between an economical model and a premium one, as shown in the table below.
| 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 balancing for most workflows |
| Claude Opus 5 | $25,00 | Lead agent or escalation for high-complexity tasks |
| OpenAI o3-pro | $80,00 | Advanced reasoning; very high cost for long workflows |
Illustrative example (the numbers are just for the math, not real customer data): an e-commerce store using Claude Sonnet 5 to review 1,000 product pages with a multi-agent workflow might use 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 out. If the process is already fast, probably not.
Three concrete scenarios: what changes for an e-commerce, a professional firm, and a manufacturing company?
Multi-agent workflows bring different advantages depending on the sector and the type of process. Here are three typical cases with an estimate of the value generated.
| Company type | Automatable process | Expected advantage | Main risk |
|---|---|---|---|
| E-commerce (200+ products) | Parallel product sheet review and optimization | Time reduction from days to hours; tone consistency | High token cost on large catalogs; human supervision is needed |
| Professional firm (legal/tax) | Document analysis on large volumes of contracts or claims | Broader coverage, fewer omission errors | Sensitive data: audit trail and EU AI Act compliance needed |
| Manufacturing company | Automated QA on technical specs and production reports | Anomaly detection across multiple lines in parallel | Integration with existing systems; initial setup cost |
To learn more about how companies are already applying these systems in real-world contexts, the article on AI agents in business workflows with three real-world case studies offers concrete adoption examples.
What tools exist today for multi-agent orchestration and how do they compare?
Claude isn't the only system that allows coordinating multiple AI agents in parallel. Alternatives exist with different features and costs: the choice depends on your existing ecosystem and the level of customization needed.
| platform | Max parallel agents | Strength | Main limitation |
|---|---|---|---|
| Claude Managed Agents (Anthropic) | Up to 1,000 (25 active threads) | Native integration, Outcomes for automatic QA | High token cost with premium models |
| OpenAI Agents SDK | Variable (depends on implementation) | Broad ecosystem, different models | A senior OpenAI engineer called 'agent swarms' a waste of tokens |
| LangGraph / LangChain | Unlimited (self-hosted) | Total flexibility and control | Requires custom development and advanced technical skills |
| Microsoft AutoGen | Variable | Integration with Azure and Microsoft tools | Steep learning curve |
The platform choice isn't just technical. If your company already uses Microsoft tools, AutoGen can integrate more easily. If you want to get started quickly without custom development, Claude Managed Agents offers a guided path. For those who want to better understand the AI agent landscape, the article on AI Agents and business workflows: what changes in 2026 is a good starting point.
What rules and risks do you need to know before enabling a multi-agent workflow?
Starting August 2, 2026, the main obligations of the EU AI Act will apply to high-risk AI systems. Multi-agent workflows that impact significant decisions — hiring, credit, security — fall into this category and require specific measures.
The EU AI Act (the European Artificial Intelligence Act) mandates three main obligations for high-risk systems, which came into force on August 2, 2026: a log of every decision made by the system (audit trail), an immediate human-manageable shutdown mechanism (the so-called 'stop button'), and a continuous risk management process. This is explained in detail Iusletter .
In a multi-agent architecture, each agent performing a high-risk function must independently comply with these obligations. It is not enough for the system as a whole to be compliant: each node in the chain must be.
| EU AI Act requirement | What this means in practice | Who is involved |
|---|---|---|
| Audit trail (Art. 12) | Log every decision made by each agent, with timestamps and reasoning | IT, Legal, DPO |
| Human oversight with stop button (Art. 14) | A human operator must be able to stop the workflow at any time | Operations, IT |
| Continuous risk management (Art. 9) | Assess and update system risks over time, not just at launch | Management, Legal |
| Compliance for each agent (Recital 99-100) | In multi-agent architectures, every high-risk agent must be compliant | IT, Legal |
If your workflow involves low-risk internal processes — like text review or analysis of aggregated, non-personal data — the obligations are less stringent. However, it's always useful to document your choices and maintain human oversight of the results.
Why are people talking about this 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 rapidly, but operational maturity is still limited.
The most important data is not about adoption, but about success: fewer than one in seven pilots makes it to production, according to Fifthrow . The factors that make the difference are clear governance, integration with existing systems, and regulatory compliance — not the choice of AI model.
For those who want to understand how other tools position themselves in this market, the 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 and well-defined use case. Here is the recommended path for an SME starting from scratch.
- 1 · Management / Operations Choose a limited use case Pick a repetitive process with parallel sub-tasks and verifiable outputs. Steer clear of high-risk processes for your first pilot.
- 2 · IT / Developers Activate the correct agent type Select 'multiagent_20261001' as the agent type in the Claude platform. Follow the official documentation or run '/claude-api managed-agents-onboard' in Claude Code.
- 3 · IT / Operations Set token and cost limits Set a maximum budget per session before starting the workflow. Multi-agent workflows can burn through tokens pretty fast.
- 4 · Internal team Manually validate output during the first few sessions Don't rely entirely on full automation right away. Use Outcomes for automated evaluation, but keep a human in the loop for critical results.
- 5 · Management / Legal Check regulatory compliance If the process touches personal data or relevant decisions, check the EU AI Act obligations before going into production.
- 6 · Management Measure and scale After the pilot, compare quality, time, and costs with the previous process. Only if the bottom line is positive, roll it out to other processes.
What are the most common mistakes people make when starting with multi-agent workflows?
Most pilots fail not due to technical problems, but due to wrong organizational choices: overly complex processes, uncontrolled budgets, and lack of human supervision.
Those new to multi-agent orchestration tend 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 is too complex | Wanting to showcase maximum potential right away | Choose a use case with verifiable output and low risk |
| Not setting spending limits on tokens | The consumption of agentic workflows is underestimated (up to 30x) | Set a maximum budget per session before starting |
| Eliminating human supervision right away | Aiming for total automation from the start | Keep a human check on results for at least the first few weeks |
| Ignoring regulatory compliance | Thinking it only applies to large companies | Check EU AI Act obligations even for internal pilots |
| Always choose the most powerful (and expensive) model | Quality is confused with model cost | Use budget-friendly models for secondary agents, keep premium ones only for the lead agent or escalation |
| Not measuring the pilot results | Moving forward out of inertia without data | Define success metrics before starting and compare them afterwards |
A useful insight on how to properly structure an agentic workflow before launching it can be found in the article on Claude Opus 4.8 and dynamic workflows for SMBs , which explores the basics of this architecture.
How to tell if the multi-agent workflow is actually working for your business?
Concrete indicators are needed, measured before and after the pilot. Output quality, cost per unit processed, and time saved are the three fundamental parameters.
| Indicator | How it is measured | Positive signal |
|---|---|---|
| output quality | Errors found or tasks completed correctly vs. previous process | Improvement ≥20% compared to the previous method |
| Cost per processed unit | Total token cost divided by the number of outputs produced | Lower than the cost of equivalent manual labor |
| Completion time | Hours or minutes to complete the end-to-end process | Significant reduction compared to the manual or sequential process |
| Residual error rate | Outputs requiring manual correction after the workflow | Under 10% of total outputs |
| regulatory compliance | Complete audit trail, working stop button, updated risk log | No documentation gaps in periodic audits |
So, is it worth your company adopting multi-agent workflows?
It is worth it if you have repetitive processes with many parallel sub-tasks (QA, document analysis, review on large volumes), a technical team capable of configuring and monitoring the system, and you are willing to invest in a pilot with clear metrics.
A middle ground your process is semi-structured and you want to automate it, but you don't have experience with AI agents yet. Start with a simple workflow with a single agent, measure the results, and then evaluate whether to add multi-agent coordination.
Not yet if your process is linear and simple, volumes are low, or you don't have the resources to manage regulatory compliance and continuous monitoring. A chatbot or a single agent is cheaper and just as effective.
How we see it at SHM
For us, multi-agent orchestration is a powerful tool — but only if used to free people from routine, not to chase technological novelty. The important data 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 tasks that truly require human judgment.
We've worked with manufacturing and consulting companies handling huge volumes of technical docs: in those cases, agent parallelization makes sense because the process is already structured and the output can be verified. In other cases — creative processes, customer relationships, strategic decisions — a single agent or partial automation is the smarter choice.
If you are evaluating where and how to introduce agentic automation in your company, our strategic AI integration always start from an analysis of real processes, not from a list of tools. For those who want to train the internal team, the path of business process automation training is designed for those without a technical background. You can also explore the topic of AI and process automation in our dedicated cluster, where you can find guides and up-to-date analyses.
Frequently Asked Questions
The stated limit is 1,000 agents per overall workflow execution, but the number of active threads simultaneously per session is 25, including the main thread. 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 chosen model: Claude Sonnet 5 costs $10 per million output tokens, and Claude Opus 5 reaches $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 involves high-risk decisions — such as personnel selection, credit access, or security — it falls under the obligations of Title III of the EU AI Act, effective August 2, 2026. In multi-agent architectures, each agent performing a high-risk function must comply independently.
According to Fifthrow, less than 15% of enterprise AI agent projects make it to production. The factors that make a difference aren't technical, but organizational: clear governance, integration with existing systems, and regulatory compliance. Starting with a limited, well-defined use case significantly increases the chances of success.
To activate Claude's dynamic workflows, you need to access the Claude Managed Agents platform and configure the correct agent type. You don't need an advanced development team for simpler use cases, but it's advisable to have at least one person with basic technical skills for configuration, cost monitoring, and error management.
A process is suitable if it can be broken down into independent sub-tasks that can be executed in parallel, if the output is objectively verifiable, and if the volumes justify the token cost. Linear, simple, or low-volume processes are better managed with a single agent or a traditional chatbot.
The real question isn't how many agents you can get running, but how many hours of repetitive work you can give back to 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.
