Liquid AI: Build Your Own Agent Framework Without Code

You spend hours copying lead data into spreadsheets, then rewriting each outreach line by hand. After that you still have to generate invoices and chase payments. By the end of this article you’ll have a liquid AI setup that does all three steps automatically, using only no‑code tools and a few prompts.

What is liquid ai and why it matters

Liquid ai describes a flow where small, specialized agents pass data to each other like water moving through a pipe. Each agent does one thing well—extract a field, write a sentence, calculate a total—and the output of one becomes the input of the next. This keeps the system easy to change and cheap to run.

For a solo operator the biggest win is that you never need to hire a developer to stitch together APIs. You can swap out one agent for another without breaking the whole chain.

How do you start building a liquid ai workflow?

Begin by picking a trigger that starts the process. A common choice is a new row added to a Google Sheet when a lead form submits. From there you add agents that act on that row.

First agent: enrich the lead with company size and tech stack using a simple web scrape prompt. Second agent: turn the enriched data into a cold‑email draft. Third agent: create an invoice PDF and email it to the client.

You can build this entirely inside a visual automation tool like Make. Make lets you connect modules without writing code, and each module works as an agent in the liquid chain.

What most guides get wrong about agent chaining

Many tutorials tell you to dump all the logic into one giant module that does enrichment, writing, and invoicing at once. That creates a monolith that is hard to test and impossible to reuse.

What works better is to keep each agent focused on a single responsibility and to pass only the data it needs. If you try to make one agent do three things you end up with a brittle flow that breaks when any input changes.

Concrete example: lead enrichment → email draft → invoice using Make, Claude, and Google Sheets

Here is the exact setup I run for my own consulting business.

• Trigger: Watch for new rows in a Google Sheet called “Leads”.
• Agent 1 – Enrichment: Use the HTTP module to call a free scrape API (like https://api.scrapingbee.com) with a prompt that returns the company’s employee count and main technology. I store the result in two new columns.
• Agent 2 – Email draft: Call the Claude API via Make’s HTTP module with the following prompt:

Write a short, friendly cold email to {{first_name}} at {{company}}. Mention that I help businesses like theirs reduce manual work by 30%. Keep it under 120 words.

The output from Claude goes into a column called “Email Draft”. I then use the Gmail module to send the email directly from the sheet.

I’ve kept the prices low: the scrape API free tier gives 1000 calls per month, Claude’s API costs $0.008 per 1k tokens, and Make’s core plan is $29/mo. For a solo operator handling fifty leads a month the total spend stays under $15.

I love how Claude can rewrite a cold email in under two seconds with a simple prompt—it saves me at least ten minutes per lead.

My gripe is that Make hides error logs behind its paid plan; on the free tier you only see a red exclamation mark with no details, which makes debugging a pain.

(Which, yes, is annoying) but the workaround is to add a simple “Set Variable” module after each step that logs the output to a separate Google Sheet for review.

How to debug when this breaks

When an agent fails, the first place to look is the module’s output log in Make. If the log is missing, check the temporary logging sheet you added.

Common failure points:

• The enrichment API returns an empty response because the scrape target blocked the request. Fix: add a retry module with a 30‑second delay.
• Claude returns a safety warning because the prompt includes a disallowed word. Fix: sanitize the input by removing any profanity before sending it to the API.
• The PDFMake module fails to generate a file because the template references a missing field. Fix: add a filter that replaces empty values with “N/A” before the PDF step.

If you still see errors, run the scenario manually with test data and watch each module’s output in real time. That usually reveals where the data shape changed.

Pricing, trade‑offs, and my honest take

Make’s $29/mo plan gives you unlimited scenarios and enough operation slots for a solo business. I think that price is fair for the time it saves.

Claude’s API is cheap but you need to monitor token usage; a long prompt can eat through credits fast. I keep my prompts under 150 tokens to stay predictable.

The free tier of Google Sheets works fine for up to a few thousand rows; beyond that you might hit the cell limit and need to move to Airtable.

We cover this in more depth elsewhere — deeper coverage of AI agent platforms.

Overall, the liquid ai approach lets you change one agent without touching the others. If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault.