There’s a temptation when we talk about AI to think in terms of a single assistant.
You give it a task. It thinks about it. It gives you an answer.
That works surprisingly well for many things.
But a lot of the work businesses do every day isn't actually one task.
Take something as familiar as applying for a grant.
On the surface, it sounds simple:
“Complete the grant application.”
But that's not really the job.
Before anyone can write anything, someone needs to find the right opportunity, understand what the funder is looking for, check whether the organisation is eligible, review the requirements, gather information, find supporting evidence, understand previous applications or projects, identify gaps, draft responses, check those responses against the criteria, review the submission for consistency and compliance, and eventually decide whether it is ready to go.
A bid process is much the same.
“Write the proposal” sounds like one task.
It isn't.
It is a chain of different tasks, involving research, analysis, interpretation, writing, checking, decision-making and sometimes several rounds of review.
And that's where I think multi-agent AI systems become particularly interesting.
One AI agent can only take you so far
The obvious way to introduce AI into a complex workflow is to give the whole thing to one AI assistant.
Give it the documents.
Give it some instructions.
Tell it what you're trying to achieve.
Then ask it to get on with the job.
For relatively simple workflows, that can work extremely well.
But as the process becomes more complicated, the limitations start to appear.
The AI has to keep track of a large amount of context.
It may be researching something at the same time as trying to write something.
It may be interpreting requirements while also deciding what information is relevant.
It may be asked to produce an answer and then review its own answer against the same criteria.
And when something goes wrong, it can be difficult to understand exactly where it went wrong.
Was the research poor?
Did it misunderstand the requirement?
Did it use the wrong information?
Did it make an assumption?
Did the drafting instruction conflict with something earlier in the conversation?
The problem isn't necessarily that the AI isn't capable.
It's that we’re asking one system to perform too many different jobs at once.
What if the work was split between specialists?
This is where the idea of multiple AI agents becomes useful.
Instead of asking one AI to manage the entire workflow, different agents can be given different responsibilities.
For example, a grant management workflow could have:
- Opportunity
- Eligibility
- Research
- Requirements
- Drafting
- Review
- Compliance
- Human Decision
An Opportunity Agent
Looks for relevant funding opportunities and filters them according to the organisation's interests, geography, sector, funding requirements and other criteria.
An Eligibility Agent
Takes a potential opportunity and checks whether the organisation appears to meet the eligibility requirements.
A Research Agent
Gathers relevant information, previous work, organisational evidence, supporting documents and other material that could be useful.
A Requirements Agent
Reads the grant guidance and breaks it down into the things the application actually needs to address.
A Drafting Agent
Uses the approved information and requirements to create an initial response.
A Review Agent
Checks whether the draft actually answers the questions being asked and whether important points have been missed.
A Compliance Agent
Checks for things such as missing information, contradictions, unsupported claims, formatting requirements and other submission rules.
And then, importantly, a human reviews the work.
The same principle can apply to bids.
You could have agents responsible for opportunity qualification, RFP analysis, requirements extraction, research, evidence gathering, drafting, review and compliance.
Suddenly, we're no longer asking AI to “write a bid”.
We're asking AI to participate in a business process.
And that's a much more interesting proposition.
The agents aren't the interesting part. The orchestration is.
It's easy to get excited about the idea of having ten AI agents working together.
But simply adding more agents doesn't make a system better.
The difficult part is deciding how they work together.
Which agent gets involved first?
What information does it receive?
What information is it allowed to access?
What does it pass to the next agent?
What happens if it isn't confident?
What happens if two agents reach different conclusions?
When does the workflow stop?
When does a human need to make a decision?
What happens if an agent fails?
These are workflow design questions as much as they are AI questions.
Imagine an eligibility agent deciding that a funding opportunity looks suitable.
That shouldn't necessarily mean the drafting agent immediately starts writing.
Perhaps the system first needs to pass the opportunity to a human for approval.
Once approved, the research agent can begin gathering evidence.
The requirements agent can analyse the application criteria.
The drafting agent can then work from the approved research and requirements.
The review agent can challenge the draft.
And if the review finds a significant problem, the workflow can send the work back for another iteration rather than simply continuing towards submission.
That's orchestration.
The intelligence isn't just in the individual agents. It's in how the agents work together.
This is different from simply adding a chatbot
There's an important distinction here.
A chatbot answers questions.
An AI assistant might help you complete a task.
A multi-agent system can become part of the process itself.
That's a meaningful shift.
Imagine someone working on a grant application.
They shouldn't necessarily have to know which AI tool to open, which prompt to use, which document to upload or which instruction to give next.
The system can manage much of that complexity behind the scenes.
The person might simply see:
Opportunity identified
Eligibility appears to be met.
Requirements analysed
14 application criteria identified.
Evidence gathered
9 relevant organisational documents found.
Draft prepared
11 of 14 criteria addressed.
Review completed
3 areas require additional information.
That's a very different experience from opening a chatbot and asking:
“Can you write my grant application?”
The AI isn't just generating content.
It's helping move the work through the organisation.
The human doesn't disappear
This is probably the most important part.
There is a temptation to describe multi-agent systems as a way of automating an entire business process.
Sometimes that's appropriate.
But in many of the workflows I'm interested in, the objective isn't to remove people.
It's to remove the work that doesn't require their judgement.
AI can research.
It can organise information.
It can compare documents.
It can identify requirements.
It can draft.
It can check.
It can highlight inconsistencies.
It can suggest improvements.
But there are still decisions that someone needs to own.
Is this actually the right opportunity for us?
Are we comfortable making this commitment?
Is this claim accurate?
Does this genuinely represent our organisation?
Is this the right thing to submit?
Those aren't simply AI problems.
They're business decisions.
And I think good multi-agent systems should make those decisions easier — not pretend they don't exist.
More agents doesn't automatically mean a better system
There's another trap here.
Once you understand the concept, it's quite easy to start designing a system with an agent for everything.
One agent researches.
Another checks the research.
Another checks the checker.
Another reviews the reviewer.
Before long, you've created an impressive architecture diagram and a workflow that nobody really understands.
That's not necessarily progress.
Every additional agent introduces another point where things can go wrong.
There are more handoffs.
More context to manage.
More opportunities for conflicting information.
More permissions to control.
More things to monitor.
And potentially more cost and latency.
So I don't think the right question is:
“How many AI agents can we use?”
The better question is:
“Where does having another agent actually improve the workflow?”
Sometimes the answer will be one agent.
Sometimes it will be several.
And sometimes a conventional piece of software or a simple automation will be better than either.
That's an important part of designing these systems properly.
The real opportunity is specialisation
For me, the most interesting thing about multi-agent AI isn't that there are multiple AIs.
It's specialisation.
We've spent decades designing software where different parts of a system have specific responsibilities.
A CRM manages customer information.
A finance system manages financial records.
A document management system manages documents.
A workflow engine manages processes.
AI now gives us another possibility.
Different AI components can specialise in different types of cognitive work.
One can be particularly good at research.
Another can analyse requirements.
Another can work with structured organisational information.
Another can draft.
Another can challenge the draft.
Another can check compliance.
And the system can bring those capabilities together around a real business workflow.
That's much closer to how organisations actually work.
People don't generally sit down and perform an entire complex business process from beginning to end without switching roles, checking information or getting input from someone else.
Why should our AI systems be designed that way?
From AI assistant to AI-enabled business process
I think this is where the conversation around AI is beginning to change.
For the last couple of years, much of the focus has been on what AI can generate.
Write an email.
Summarise a document.
Create an image.
Write some code.
Draft a proposal.
Those capabilities are useful.
But the bigger opportunity may be what happens when AI becomes part of the workflow surrounding those tasks.
A grant application isn't just writing.
A bid isn't just writing.
Customer support isn't just answering.
Sales qualification isn't just asking questions.
Document processing isn't just extracting text.
Each of these processes contains a sequence of activities, decisions, checks and handoffs.
That's where AI starts becoming much more interesting.
Not as a clever chatbot sitting beside the business.
But as an intelligent layer inside the business process.
And that's probably where the hard work begins
Building a multi-agent system isn't simply a matter of connecting several models together.
You need to think about the information each agent can access.
You need clear responsibilities.
You need reliable sources of truth.
You need to control what agents are allowed to do.
You need to know when an answer is based on evidence and when it is an inference.
You need human approval at the right points.
You need logging and auditability.
You need to test individual agents as well as the overall workflow.
And when the system goes into production, you need to monitor it.
Because a demo that works beautifully for ten examples isn't the same thing as a system that an organisation can depend on every day.
That's the part of AI implementation that I find particularly interesting.
The AI model is only one part of the system.
The real work is designing everything around it so that the system is useful, controlled and dependable.
So, when does multi-agent AI make sense?
I wouldn't recommend it simply because it sounds sophisticated.
I'd look for a few characteristics.
The process involves multiple distinct stages.
Different stages require different types of reasoning or information.
There are meaningful handoffs between activities.
The process contains a lot of research, comparison, drafting and checking.
There are clear points where human judgement is required.
And, importantly, the process happens often enough that improving it would create meaningful business value.
Grant management and bid management are good examples because they contain many of these characteristics.
But the same thinking can apply elsewhere.
Customer operations.
Compliance.
Procurement.
Sales operations.
Internal knowledge.
Professional services.
Document-heavy processes.
The opportunity isn't limited to any particular industry.
The question I keep coming back to
We're still figuring out what these systems should look like.
The technology is moving quickly, and some of the terminology will probably change along the way.
But I think the underlying idea is going to stick.
AI doesn't have to be a single assistant that tries to do everything.
It can become a collection of specialised capabilities working together, with software orchestrating the workflow and people providing judgement where it matters.
And that changes the question we should be asking.
Not:
“What can AI generate for us?”
But:
“Which parts of this process should AI handle, which should people handle, and how should the two work together?”
That's a much more useful question.
Because the goal isn't to add more AI.
It's to redesign the work.
And when the right AI is doing the right job at the right point in a workflow, the result isn't simply a better AI assistant.
It can be a fundamentally better way of getting the work done.
Author
Vikram Katyani — Founder, IntelliMinds Digital.