How much Agency should your Automation get?

Last year at Singapore’s FinTech Festival, I was talking to investors about a private market AI platform I was helping to build at the time. More than one of them asked me some version of the same question: “Why do I need your platform if Claude or ChatGPT can already do this for me?”

It’s a fair question. The meat of the platform was to pull data from a few sources, organise them into a schema, work out comparison metrics, then do the same for the next company. An agent with the right tools could do all of that.

Not long ago, that list was a workflow you’d have to build by hand, one step at a time. Now agents have gotten so good that with a few tools and a well-crafted prompt, they work out the steps themselves. Agents seem to have replaced these automated workflows.

Some introspection led me to the core assumption of the question. It assumes the task runs once. Most automation doesn’t. It runs every day, for every new customer, for every company on the list. On every run, the agent works out the same plan it worked out last time. It pays for the same reasoning again, and it doesn’t always reach the same decisions. The table for company 1 and the table for company 50 might not be filled in the same way. Doing it once is a proof of concept. Doing it with consistency and reliability over a thousand times is what makes it a product.

Agents won the first run. The thousandth run is where workflows come back, and where the real question starts: how much agency should your automation get?

Run One

Agents can easily one-shot a workflow for you. And in fact, they’re great at it. Take the AI platform for Investors as an example. Add connections to data platforms like PitchBook and various search engines, ask for a due diligence report on a company, and voilà.

So how does an agent pull that off on the first try? Three things.

1. It fills in the gaps: If you’ve missed out on a format to use, an aspect to consider, or a source to look at, the agent will make these decisions for you.

2. It decides from what it sees: If the agent hits a snag in one of its sub-tasks, like not being able to find a key piece of information from a primary data source, it’d react to that situation and start looking elsewhere to fulfil the data point.

3. It starts from scratch: It doesn’t need to have seen the task before, because it plans the whole thing as it goes.

So here’s what it looks like under the hood: You ask for a report on a company. The agent searches PitchBook and finds its funding rounds, but no revenue. So it searches again, and finds a revenue figure for the last fiscal year in a press release. It fits what it found into a table, works out a few comparison metrics, and then writes the report.

ONE AGENT · ONE REPORT · RUN ONE

Your prompt

Task

A due diligence report on Company A

Tools

PitchBook, web search

Format

not givenOverview · Funding · Financials · Comparables · Risks

Actions it can take

• 1Search PitchBook
• 2Search the web
• Search the news
• Read a filing
• 3Fill in the table
• 4Work out metrics
• 5Write the report

The run

1. Pick a formatThe prompt doesn't give one, so it picks five sections: overview, funding, financials, comparables, risks.Fills in the gaps
2. Plan the stepsPicks 4 of the 7 actions it can take, and puts them in order.Starts from scratch
3. Search PitchBookFunding rounds and a $300M valuation. No revenue.
4. Search the webNot in the plan: PitchBook had no revenue, so it looks elsewhere. A press release gives $42M for fiscal 2025.Decides from what it sees
5. Fill in the tableRevenue: $42M, from the press release.
6. Work out metricsValuation / revenue: 7.1x
7. Write the reportFive sections, in the format it picked at step 1.

All you really have to do is give the agent the tools you want to connect, describe how you want them connected, and state the end goal. The agent will handle the rest. Even better if you can provide context and key considerations in the prompt.

Run One Thousand

The issue with using agents becomes clearer after running them many, many more times. The 2 problems are cost and drift.

While great that your agent can start the workflow from scratch, the fact that it starts from scratch at every run means you’re likely paying for reasoning tokens that led to decisions it had already made before. Nothing it worked out for company 1 carries over to company 2. Prices per token keep falling, but an agent still pays for the same reasoning on every run.

In terms of drift, it happens because agents have agency to make decisions for themselves. They help to fill in the gaps and make decisions to patch blockers. What this means is that they can guarantee that your data table will have all its fields filled with relevant numbers. But each number could very much come from different sources, and these are solely based on the agent’s judgement.

THE SAME TABLE · FOUR COMPANIES

An agent, one run per company

Company	Revenue	Source	Window202420252026
Company A	$42M	press release	fiscal 2025202420252026
Company B	$55M	news article	run-rate: Aug 2026 × 12202420252026
Company C	$38M	analyst estimate	last 12 months202420252026
Company D	$61M	annual report	fiscal 2024202420252026

Every field filled. Revenue from 4 sources, covering 4 windows.

A fixed workflow

Company	Revenue	Source	Window202420252026
Company A	$40M	annual report	fiscal 2025202420252026
Company B	$47M	annual report	fiscal 2025202420252026
Company C	$35M	annual report	fiscal 2025202420252026
Company D	$66M	annual report	fiscal 2025202420252026

Every field filled. Revenue from 1 source, covering 1 window.

For a scenario like this, where we are comparing companies, it’s a deal-breaker. Reliability, having numbers obtained the same way, ideally from the same source, is what makes the outputs useful at all. An agent that works the task out from scratch can’t ensure that. So agents win for single runs. But anything more would require something more reliable. Like a fixed workflow.

The New Age of Workflows

So, workflows are back. Sort of. The first run still belongs to the agent. Nobody wants to go back to wiring every step by hand. So the question in the new age of automation isn’t agent or workflow. But how much agency a task should get: how many of its decisions you leave to an agent, and how many you decide in advance.

The way I see it, there are 2 ways to look at this:

1. Agent first, workflow second: The agent stays in charge, but works with a fixed set of pre-defined workflows and tools instead of planning every step itself. It decides which ones to run and when.
2. Workflow first, agent second: The workflow stays in charge. And agents only step in at specific points. Each with a small, well-scoped task, with context it needs to exercise judgement.

ONE TASK · WHO'S IN CHARGE?

Agent first, workflow second

The agent picks which workflows to run, and in what order.

"How does its revenue growth compare with its peers?"

Agent

Workflows it can call

• Pull PitchBook data
• 2Pull annual reports
• 1Find competitors
• Map fields to schema
• 3Work out metrics
• Write report

Workflow first, agent second

The workflow sets the order. The agent only decides inside its step.

Company 1 of 200

1. 1Pull PitchBook data
2. 2Pull annual reports
3. 3Read pitch deckagent
4. 4Pick sectorJev
5. 5Map fields to schema
6. 6Work out metrics

Inside steps 3 and 4

3Read pitch deckagent

Gets

The deck (22 slides), and three fields to fill.

Returns

Product

Payments API for online marketplaces

Customers

1,200 marketplaces

Business model

Fee per transaction

4Pick sectorJev

Gets

The fields from step 3, and a fixed list of sectors.

Returns how sure it is of each

• Fintech0.94
• Healthtech0.01
• Climate0.00
• Enterprise software0.04
• Consumer0.01

agent: judgement decision model fixed: code

Agent First, Workflow Second

Think of a pilot. The pilot decides what the flight needs, but doesn’t improvise an engine restart. There’s a checklist for that, and it runs the same way every time. The agent is the pilot, and the workflows are its checklists.

The tools for this are already here. MCP (the Model Context Protocol) gives every tool the same interface, so any agent can plug into it. Skills package a task as instructions and scripts that the agent loads only when it needs them. Anthropic’s Agent Skills post says why that matters: “because code is deterministic, this workflow is consistent and repeatable.” Automation platforms followed. Workato’s genies are agents whose skills are workflows, and n8n can expose any workflow to an outside agent as an MCP tool.

Where it shines is work you can’t plan in advance. Say the report on a company is done, and the investor starts asking follow-ups. “How does its revenue growth compare with its peers?” “Which of its competitors raised money this year?” Each question needs different steps, so there’s no fixed sequence to build. The agent picks which workflows to run, and the numbers they return come out the same way every time. Workflow first would need a branch for every question nobody has asked yet.

Workflow First, Agent Second

Think of an assembly line. The line sets the order of the work, and people stand only at the stations that need a judgement call, like spotting defects. The workflow is the line, and the agents are the people at its stations.

This is how automation platforms already work, with agents added as steps. n8n’s AI Agent node sits inside a workflow like any other step. In Workato, a workflow can hand a task to a genie, pause until it’s done, and carry on with its answer in fields the next step can use.

Where it shines is the same task, run on every item of a list. Say 200 companies come through the pipeline every month, and each one gets the same comparison table. Pulling funding rounds from PitchBook, mapping fields into the schema and working out the metrics are the same for every company, so they’re plain code. An agent steps in only to read the company’s pitch deck, which is laid out differently every time. It gets the deck and the fields to fill, nothing more. Picking the company’s sector from a fixed list doesn’t even need an agent. A decision model like Jev can make that call and say how sure it is.

Agent first wouldn’t fit, even with the same workflows as tools. The agent still decides which ones to run, 200 times a month, and it won’t always decide the same way. That’s run one thousand all over again.

Best of both worlds?

“So which one should I use?”

Neither, on its own. Most automations sit somewhere in between, and the line moves as you learn where the judgement really is. A step the agent always decides the same way can become a fixed step. A fixed step that keeps breaking on new inputs can go back to an agent. You don’t even have to write the fixed steps yourself. After a run, the agent can turn the decisions it made into a script, and the next run calls the script instead. As n8n’s Production AI Playbook puts it: “Use AI only where you need AI.”

ONE PIPELINE · HOW THE LINE MOVES

Finding revenue and writing the report settled too. Mapping fields went back to the agent.

1. Find sourcesfixed
2. Pull funding roundsfixed
3. Find revenuefixednow fixedSince run 30: the annual report, last fiscal year.
4. Map fields to schemaagentback to the agentRun 57: a report in a new layout broke the script.
5. Read pitch deckagent
6. Work out metricsfixed
7. Write reportfixednow fixedEvery report used the same five sections.

Which brings us back to the opening question: what’s the platform for? Well, it’s for creating, managing, and governing these workflows.

Imagine one that saves versions of your workflows, shows you which steps to break down further or replace with a scoped agent, lets you plug workflows into other workflows, and manages the agents, LLM calls and MCP servers they all use. You don’t need any of it to run a workflow a thousand times. But a workflow that runs a thousand times is still just a workflow. The rest is what makes it a product. You build faster, every change leaves an audit trail, and every workflow and its building blocks stay governed. That’s what lets a business trust it with their own work.