Multi-Agent Workflows: Getting Several AI Agents to Work Together

When people discover AI agents, they're tempted to give everything to a single one: “find prospects, write to them, follow up and send me a report.” It can work, but the agent quickly becomes hard to tune. If the follow-up is clumsy, do you fix the research, the writing or the tracking? If one step fails, everything stops.

In a company, you don't hand an entire process to one person without some organization. You split the roles, and each person passes their work to the next. Multi-agent workflows apply the same logic: several specialized agents, linked together, each with a clear mission.

What is a multi-agent workflow?#

A multi-agent workflow is a sequence of AI agents where the output of one becomes the starting point of the next. Each agent keeps its own brief, tools, trigger and log. What connects them is:

• shared data: usually a table that the first agent fills and the next one reads;
• a chained trigger: the downstream agent starts when the upstream agent has finished successfully.

Why several agents rather than one#

The main benefit is clarity. You know who does what, you see what passes between steps, and you can improve one step without touching the others.

How it works on Future Agent#

Shared tables#

An agent can fill a shared table: a kind of spreadsheet internal to your account, with defined columns (date, supplier, amount, status…). Each row is identified by a key, which prevents duplicates when the agent goes over the same data again. These tables can be exported and count toward your storage.

Chained triggers#

An agent can be linked to an upstream agent. When the upstream agent completes a successful run, the downstream agent starts automatically. If the upstream run fails, nothing is passed on: errors don't propagate.

Only new rows passed along#

The downstream agent only receives rows added since its last run. It doesn't reprocess the whole table every time. If the collector adds three invoices tonight, the reconciler only works on those three.

Memory as a relay#

When searching its memory, a downstream agent can find relevant items left by the agents before it. Context travels along with the data. See our article on AI agent memory.

The workflow map#

The dashboard offers a 3D map of your agents and their links: you can see which agents feed which tables, which ones are linked and which ones are currently working.

Three example chains#

Invoices, reconciliation and accountant#

1. The collector, Monday to Friday at 7 p.m., spots invoices received as attachments, files the PDFs in a shared folder and adds one row per invoice to the “Invoices” table (date, supplier, amount excluding tax, VAT, file).
2. The reconciler starts after the collector. For each new invoice, it finds the matching payment and flags discrepancies: an invoice with no payment, an amount that doesn't match.
3. The archivist, on the 5th of the month, sends your accountant the previous month's package with the Excel summary and the list of discrepancies.

Template: Invoices, reconciliation and accountant.

End-to-end prospecting#

1. The prospector finds companies that match your criteria and adds them to a prospects table.
2. The outreach agent drafts a first personalized message for each new prospect.
3. The follow-up agent chases those who haven't replied after the delay you set.

Template: Prospecting, outreach and follow-up.

Monitoring, writing and publishing#

1. The monitoring agent spots trending topics in your industry.
2. The writer drafts an article from the selected topics.
3. The publisher turns each article into a LinkedIn post, submitted for your approval before it goes live.

Template: Monitoring, writing and publishing.

The library offers fifteen ready-to-adapt chains, including industry-specific ones: quotes, scheduling and invoicing for construction (Quote, job site and invoice), reminders and waitlist for a medical practice, cover forecasts and orders for a restaurant.

Building your first chain, step by step#

Step 1: Sketch the process on paper#

Write down the steps of your current process, in order, with what passes from one step to the next. Example:

1. Receive quote requests → list of requests.
2. Price each request → quotes sent.
3. Follow up on unanswered quotes → quotes followed up.
4. Review every Monday → report.

Step 2: Identify the data that flows#

For each arrow, what information needs to pass? These will be the columns of your shared table: client, date, amount, status, link to the document. Also choose what uniquely identifies a row (a quote number, for example).

Step 3: Set the rhythm of each step#

Not every step needs the same rhythm. Collection can run every evening, processing right after, the report once a week. Only steps that must immediately follow the previous one are chained; the others keep their own schedule.

Step 4: Create the first agent and let it run#

Start with the upstream agent. Let it fill its table for a few days and check the quality of the rows: columns properly filled, no duplicates, consistent formats. A chain is never better than its first step.

Step 5: Add the next agent#

Describe the second agent, specifying that it works from the first one's results. On Future Agent, the architect knows your existing agents: it can suggest linking the new agent to the one that feeds the relevant table.

Sample brief for a downstream agent:

After each run of my “Request collection” agent, take the new requests from the “Quote requests” table. For each one, prepare a quote based on my price list (attached context file), clearly stating the assumptions. Save the quote as a PDF, add its number and amount to the table, and send it to me before anything goes out.

Step 6: Watch the whole chain#

During the first few weeks, follow one item from end to end: an invoice, a prospect, a request. Check that it moves correctly through each step. Each agent's log and the workflow map show you where it is.

Mistakes to avoid#

• Too many agents too soon. Start with two linked agents; add the third once the first two are stable.
• Fuzzy columns. “Info” or “Misc” columns cause problems later. One column, one piece of information.
• No unique key. Without an identifier, duplicates pile up when an agent goes over the same data again.
• Loops. Don't let a downstream agent trigger the upstream agent in turn: the process needs a beginning and an end.
• Forgetting human approval on the step that touches your clients or your accounting.

Evolving an existing chain#

A chain isn't set in stone. Three changes come up often:

• adding a step at the end: a monthly review agent that reads the same table as the others, without changing anything that already exists;
• plugging in a second downstream flow: the requests table feeds both the agent that prices quotes and an agent that alerts the regional salesperson;
• replacing a step: if the writing agent no longer suits you, you fix or recreate it without touching collection.

In every case, keep the same discipline as at the start: one change at a time, and one item followed end to end to check that everything flows as planned.

When you don't need a chain#

A multi-agent workflow only makes sense if the process has several distinct steps, with different rhythms or tools. For a simple task (triaging the inbox, sending a morning brief), one agent is enough. For a list of simple task ideas, see 25 tasks to hand off to an AI agent.

What about pricing?#

Each agent in a chain has its own monthly price, quoted before launch. A chain of three simple agents can therefore fit within a plan that includes several agents. Details of the Free, Pro and Studio plans, and what each includes for workflows and storage, are on the Pricing page.

In short#

• A multi-agent workflow links specialized agents: each handles one step and passes its work to the next.
• The link is a shared table plus a chained trigger after each successful run.
• The downstream agent only receives new rows, and an upstream error doesn't propagate.
• Start by sketching the process, define the columns, then add agents one by one.
• The library offers fifteen ready-to-adapt chains.

To go further, compare this approach with workflow tools in our AI agent vs Zapier, Make and n8n comparison, or explore all the features.