Why This AI Development Matters to Your Business

If your team is small, every repeated task competes with customer service, sales follow-ups, operations and compliance work. You may already use several software tools, but someone still has to read documents, check information, prepare drafts, move data between systems and confirm that the result is accurate.

That is where newer multi-agent AI systems may become useful. Instead of sending every request to one powerful model, they can choose different models for different parts of a job. A simple lookup may go to a lightweight model, while a complex research or coding task may be handled by a stronger one.

Sakana AI’s new Fugu Max and Fugu Ultra v2 show how this approach is developing. Fugu is not one foundation model. It is an orchestration system that routes work across a pool of models through one application programming interface (API). The release is aimed at two business needs: stronger results on difficult work and better efficiency on routine work.

TL;DR

Multi-agent orchestration means software decides which AI model should handle each part of a task. For your SME, this could support document processing, customer enquiries, reporting and internal automation without using the most powerful model for every request.

Fugu Max focuses on output per dollar, while Fugu Ultra v2 targets complex reasoning, research and software engineering. Treat benchmark claims as signals, then test the system against your own Malaysian business workflows.

What This Means

A normal AI setup often looks simple: your application sends a prompt to one model, receives an answer and displays it to a user. This can work well, but the same model may handle everything from classifying an email to analysing a long contract.

Multi-agent orchestration adds a decision-making layer. The orchestrator reads the task, breaks it into steps and assigns suitable roles. One model may act as a “Thinker” to plan the work, another as a “Worker” to complete a step, and another as a “Verifier” to check the output. Sakana describes its system as building an agentic scaffold for each query rather than following one fixed workflow.

This matters because business tasks are not equally difficult. Extracting a delivery date from a purchase order is different from comparing several supplier proposals. Drafting a standard reply is different from investigating an unusual customer complaint involving multiple documents.

The practical idea is simple: use enough AI capability for the job, not the maximum capability for every job.

Fugu Max is designed to widen the pool of models available to the orchestrator, including open-weight and specialised models. Sakana reports that it achieved the best overall result on six listed benchmarks and expanded the cost-performance Pareto frontier on seven of ten benchmarks. These are vendor-reported results, so you should validate them with your own data before making a production decision. Source: MarkTechPost

Fugu Ultra v2 is aimed at harder tasks such as autonomous research, visual reasoning and full-stack software development. Sakana reports a Chartography score of 48.3 and a DeepSWE score of 74.3. The figures may be useful for comparing technical direction, but they do not automatically show how well the system will perform with your inventory files, customer messages or accounting documents. Source: MarkTechPost

How This Applies to Malaysian SMEs

1. Customer service and WhatsApp enquiries. Many Malaysian businesses receive questions through WhatsApp, email, social media and website forms. A routing system could classify each message first: product question, delivery status, warranty request, quotation request or complaint. A lightweight model could handle common questions using your approved knowledge base, while a stronger model reviews unusual or sensitive cases. Your staff can then focus on conversations that genuinely need human judgement.

For a local retailer, distributor or service company, the workflow could also detect language preference. A customer may write in Bahasa Malaysia, English or a mixture of both. The system can prepare a draft in the same language, identify missing information and send the case to a person when the request involves refunds, legal concerns or an upset customer. You should keep approval controls in place rather than allowing an AI agent to make commitments automatically.

2. Purchase orders, invoices and delivery documents. An orchestrated workflow can read a purchase order, extract the item codes and quantities, compare them with an invoice, then flag mismatches for your operations team. One model may perform optical character recognition and extraction, another may check the totals or dates, and a verifier may identify uncertainty. This is particularly useful when suppliers send documents in different layouts.

You still need clear rules. The system should not silently approve a discrepancy, change a supplier record or release an order without confirmation. Start with a “review and suggest” process. Measure how often the system correctly identifies missing fields, duplicate documents and quantity differences before expanding its permissions.

3. Sales proposals and follow-up. A small sales team may lose time searching old proposals, product information and meeting notes. A multi-agent workflow could summarise a customer conversation, identify requested features, prepare a proposal outline and suggest follow-up questions. A separate verifier can check whether the draft uses current product information and approved claims.

For Malaysian SMEs selling to different industries, this approach can help create more relevant drafts without requiring every salesperson to become a prompt expert. Your team should still review technical specifications, delivery commitments and regulatory statements before sending anything to a customer.

4. Internal reporting and management decisions. You could ask an orchestrated system to combine sales data, outstanding tasks and service issues into a weekly management brief. One agent can gather information, another can identify changes and a third can challenge unsupported conclusions. This is more useful than simply asking one chatbot to “analyse the business” because each step has a defined responsibility.

However, connect only the data that is necessary. Keep access rights aligned with job roles, remove unnecessary personal information and maintain a record of the sources used in every report. For a small company, a simple spreadsheet export may be a safer pilot than a direct connection to every business system.

Fugu release	Primary focus	Reported indicator	Possible SME use
Fugu Max	Efficiency and output per dollar	Best overall score on 6 listed benchmarks	Classification, document routing and routine support
Fugu Max	Broader model selection	Expanded the frontier on 7 of 10 benchmarks	Matching simple tasks to lighter models
Fugu Ultra v2	Complex reasoning and research	48.3 on Chartography	Visual document and structured-data analysis
Fugu Ultra v2	Software engineering	74.3 on DeepSWE	Reviewing code, tests and technical workflows

All benchmark figures in this table are reported by Sakana AI through the source article and should be independently tested for your use case. Source: MarkTechPost

Practical Takeaways

• Start with one workflow: choose a repeated process such as enquiry classification, document checking or weekly reporting.
• Separate easy and difficult tasks: use a lighter process for standard requests and send exceptions to a stronger model or staff member.
• Add a verifier: require a second check for customer-facing content, figures, dates and compliance-related information.
• Keep humans in control: do not allow the system to approve refunds, alter master records or make binding commitments without review.
• Track useful measures: record processing time, correction frequency, escalation volume and staff acceptance of the outputs.
• Check data handling: confirm where information is processed, how access is controlled and whether the service is available for your operating location.
• Test with Malaysian examples: use Bahasa Malaysia, English, mixed-language messages, local document formats and your actual product terminology.
• Avoid vendor lock-in: prefer systems that can work with more than one model or provide a practical export and replacement path.

The Bigger Picture

The important change is not simply that another AI model has been released. The larger trend is toward AI systems that manage other models and tools. For an SME, that could make automation more flexible: routine work can follow a fast path, while complicated cases receive additional analysis.

It also changes how you should evaluate AI. Instead of asking, “Which model is the smartest?” ask, “Which process can reliably decide what needs to happen next?” A good system should show its steps, identify uncertainty, protect sensitive data and provide a clear handover to your team.

Availability and governance still matter. Fugu Max and Fugu Ultra v2 are offered as hosted APIs, with an OpenAI-compatible interface, but the source article states that they are not available in the EU/EEA and do not have open weights for self-hosting. You should therefore confirm service availability, data residency, contractual terms and business continuity before connecting operational information. Source: MarkTechPost

For your business, the sensible next step is not to deploy a swarm of agents everywhere. Pick one frustrating process, map the decisions, test a controlled pilot and make the system prove its reliability. If it can reduce repeated work while keeping your staff responsible for important decisions, multi-agent orchestration may become a practical part of your automation roadmap.

Ready to Streamline Your Operations?

Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →