Agentic workflows are AI-driven processes in which agents plan, use tools, make decisions and adapt their actions to achieve a goal. Executing them reliably inside a business takes more than a capable AI agent: it takes a way to start work automatically, keep progress over time, connect business tools, handle approvals and track every step. That is the job of an execution harness.
An agentic workflow is a goal-driven process run by AI. One or more AI agents work out the steps, use tools to take them, check the results and change course when something unexpected happens, until the goal is met. The idea is broad: Google's examples include troubleshooting and modernising software, and the same pattern fits business processes such as chasing an invoice or qualifying a lead. Askpilot is built for that kind, agentic business workflows.
Definition
Agentic workflow: a process in which one or more AI agents plan, act through tools and adapt to what happens, to reach a goal with little step-by-step instruction from a person.
Google Cloud describes agentic workflows as AI-driven processes in which autonomous agents reason, plan and use external tools to complete complex, multi-step tasks, instead of following rigid, pre-defined paths (Google Cloud, What are agentic workflows?). The UK Government Digital Service puts it in business terms: AI agents that manage, coordinate and execute tasks within business processes, and adapt to changes and unexpected events as they go (GOV.UK, AI Insights: Agentic workflow).
Four traits come up in every serious definition:
It starts from a goal
The agent works towards an outcome, such as getting an invoice paid, rather than running a script.
It plans
The agent breaks the goal into smaller tasks and decides their order. GOV.UK calls this dynamic task decomposition.
It acts through tools
It reads and writes in other systems. For a business, that means email, WhatsApp, the CRM, the accounting software or the phone system.
It adapts and remembers
It checks what happened, corrects itself, and keeps track of how far it has got, so the work can carry on later.
Traditional vs agentic
Agentic workflows vs traditional automation
Most automation in businesses today follows steps drawn in advance. GOV.UK describes legacy workflow systems as built on "fixed, sequential steps", and Google contrasts agentic workflows with rigid automation scripts on one side and chatbots that answer one prompt at a time on the other. Agentic workflows sit between the two: they pursue a goal on their own, but within rules you set.
Traditional automation
Chatbot or AI assistant
Agentic workflow
What starts it
A trigger runs a fixed sequence
You type a prompt
A real event or a request, with a goal
Who decides the steps
Whoever drew the flowchart
You, one prompt at a time
The agent, within your rules
When the unexpected happens
It stops, or takes the wrong branch
It answers if you ask
It adapts, or asks a person
How long it works
Seconds, one run
While the chat is open
As long as the case needs: days or weeks
How you change it
Rewire the flowchart
Write a new prompt
Edit a sentence in the instructions
Best for
Identical, high-volume steps
Questions and one-off drafts
Work that varies, waits and needs judgement
Neither older model disappears. GOV.UK expects traditional and agentic workflows to run side by side, and that matches what we see in practice: fixed automation stays the right tool for identical, high-volume steps, while agentic workflows take the work that varies, waits and needs judgement. We wrote more about that line in What is agentic workflow automation?
The loop
How an agentic workflow runs
Whatever the task, an agentic workflow moves through the same loop. GOV.UK describes its lifecycle as planning, execution, iteration and monitoring. Google names the patterns behind it: planning, tool use and reflection, the agent checking its own results and correcting them.
1
Plan
The agent reads the goal and the rules, and turns them into tasks.
2
Act
It carries out the next task with the tools it has: it sends an email, updates a record, reads a reply.
3
Check and adapt
It looks at the result. If a reply changes the picture, it re-plans: a new task, a different channel, or a question for a person.
4
Remember and report
It records what it did and where it got to, so the work can continue later and anyone can see what happened.
The loop repeats until the goal is met. In a business that can take minutes or weeks: an invoice gets paid after the third reminder, a lead books a call a week after the first message.
Execution harness
What is agentic workflow execution?
A capable AI agent can plan and act, and some agents can already keep working in the background. Running a business process asks for more. Work has to start by itself when something happens, a case can stay open for days or weeks, the agent needs the company's tools and rules, some actions need a person's approval, and every step needs a record. Agentic workflow execution is running agentic workflows that way, and the software environment that makes it possible is an execution harness.
Definition
Agentic workflow execution: running agentic workflows reliably in production, from the event that starts a case to the moment it is done, with the tools, memory, approvals and record of every step that the agent needs around it.
The model does the thinking. The harness is the software environment around it that supplies state, tools, execution logic and constraints, so the work runs reliably. LangChain draws a similar line: a framework gives you building blocks, a runtime gives you production infrastructure such as durable execution, and a harness comes "batteries included" (LangChain, Agent Frameworks, Runtimes, and Harnesses, October 2025). Coding agents made the idea familiar: Claude Code is a harness for writing software. Askpilot provides a purpose-built execution harness for agentic business workflows, made of six parts:
Triggers
Real events start the work: an invoice goes past due, a lead emails, a call ends.
Sessions that last
Each case gets its own session, which can wait days for a reply and pick up where it left off.
Tools
The agent works in the systems you already use, through integrations and MCP servers.
Instructions
A plain-language PROCESS.md says what to do, what never to do and when to ask.
Guardrails
Actions you mark wait for a person's approval, and the agent is told to ask rather than guess when a fact is missing.
A record of every step
Every event, action and decision is kept, so you can see what happened and why.
These match what both sources say agentic systems need. Google lists tools, memory that lasts across sessions, and a person's approval before high-stakes actions. GOV.UK points to an audit trail for transparency, guardrails against unintended behaviour, and memory of how far along the process the system is.
How Askpilot executes it
How Askpilot executes agentic workflows
Askpilot is an agentic workflow execution platform. It has its own AI agent, the Ask agent, which executes every workflow. You can set a workflow up in the Askpilot app, or from ChatGPT, Claude or any AI agent you already use, and Askpilot runs it from start to finish.
01
Set up the workflow
Create it in the Askpilot app, or describe it to ChatGPT, Claude or any AI agent you already use, for example "chase unpaid invoices until they are paid". With Askpilot MCP connected, your AI agent writes the workflow as a PROCESS.md and puts it live.
02
An event starts a session
When an invoice goes past due, Askpilot opens a session for that invoice.
03
The Ask agent plans the tasks
Askpilot's own agent, the Ask agent, reads the PROCESS.md and turns its todos into tasks for this case.
04
It works across your tools
It checks what is owed, sends the reminder, and listens for a reply or a payment.
05
It waits, follows up and asks
When the customer replies or calls, the session wakes up. When an action needs approval, such as offering a payment plan, it stops and asks you.
06
You see every step, and change it by asking
Each session shows its progress and its full history. To change the workflow, edit it in the app or ask your AI agent to update the PROCESS.md.
Invoice chasing progress
Invoice past due
Check what is owed and how late it is
Send the first reminder
Chase every three working days until it is paid
Offer four monthly payments of £620Human approval required
Send the agreed plan in writing
2 of 5 tasks done
Each session's progress in Askpilot: tasks open and close, one waits for approval, then the next workflow takes over.
PROCESS.md
PROCESS.md, Askpilot's format for agentic workflows
Traditional automation is mapped as a flowchart, with every box and branch drawn in advance. That suits fixed steps, but it has no place for judgement, and it breaks on the case nobody drew. A prompt is the opposite: flexible, but loose, with no agreed place for the rules, the approvals or the events to wait for.
We created PROCESS.md to sit between the two. It is Askpilot's format: a plain-language Markdown file with a fixed structure, so an agent can execute it and a person can read it, own it and change it. The format is published openly, so any AI agent can write one, and Askpilot runs it.
Traditional automation
Moved into PROCESS.md format
Agentic workflows (PROCESS.md)
Invoice chasing
Chase overdue invoices until they are paid.
Invoice past due
invoice-chasing / PROCESS.md
## Role
Act as our accounts assistant, writing to customers.
## Instructions
Polite at first, then firmer. Quote amounts from the invoice. Never guess.
## Events
Subscribe to events to listen for a reply or a payment.
## Rules
Stop: the invoice is paid or written off.
Escalate: after five reminders with no reply, to the account owner.
Approval: required before offering a payment plan.
## Todos
Check what is owed and how late it is
Send the first reminder
Chase every three working days until it is paid
## Paths
wrong-amount If they say the amount is wrong, check the invoice and send a corrected copy.
The same invoice chasing workflow, drawn as a flowchart and written as a PROCESS.md.
Every PROCESS.md has the same sections:
Rolewho the agent acts as, and who it is talking to.
Instructionshow to do the work, in your own words.
Eventswhat to listen for while a case is open, such as a reply or a payment.
Ruleswhen to stop, when to escalate, and what needs approval. Rules take priority over the rest of the file.
Todosthe steps the agent plans its tasks from, in order.
Pathsextra steps for situations that need a different route.
Here is the invoice chasing workflow as a file, invoice-chasing/PROCESS.md:
invoice-chasing/PROCESS.md
---name: invoice-chasingdescription: Chase overdue invoices until they are paid.---## Role
Act as our accounts assistant, writing to customers.
## Instructions
Polite at first, then firmer. Quote amounts from the invoice. Never guess.
If they say the amount is wrong, run path wrong-amount.
## Events
Subscribe to events to listen for a reply or a payment.
## RulesStop: the invoice is paid or written off.
Escalate: after five reminders with no reply, email the account owner.
Approval: required before offering a payment plan.
## Todos
[ ] Check what is owed and how late it is
[ ] Send the first reminder
Description: Draft from assets/reminder-template.md and include the payment link.
[ ] Chase every three working days until it is paid
## Paths### wrong-amount
[ ] Check the invoice and send a corrected copy
Nobody has to write this by hand: describe the workflow to your AI agent and it writes the file. The full specification is in the PROCESS.md guide and the docs.
Examples
Examples of agentic workflows
Agentic workflows suit work that looks like a case file rather than a pipeline: something happens, and the work goes on until it is resolved.
Invoice chasing
Starts when an invoice goes past due.
Checks what is owed, reminds, chases every three working days and asks before offering a payment plan.
Lead qualification
Starts when a new lead emails.
Replies within minutes with your questions, scores the answers and chases on WhatsApp until a call is booked.
Confirms the arrears, sends the notice and books an adviser if the client mentions hardship.
Document chase
Starts when a new client signs.
Asks for each document, checks what arrives and chases what is still missing.
Guardrails
Keeping agentic workflows safe
Autonomy is only useful when it is bounded. Both sources are clear about this. Google calls a person's approval a mandatory control before high-stakes or irreversible actions, such as financial transactions or sensitive messages to people outside the business. GOV.UK names three risks to design for: a lack of transparency, unintended behaviour, and errors in the underlying model. These are the controls Askpilot uses:
People approve what matters
Mark an action as needing approval in the PROCESS.md, and the agent stops and shows you what it is about to do before it does it.
Asking instead of guessing
The Ask agent is instructed to ask for a missing fact rather than invent one, and to stop when it is unsure.
Rules take priority
The rules in a PROCESS.md rank above its other instructions, so a stop or approval rule applies even when the rest of the file points elsewhere.
Every step is on record
Each session keeps its events, actions and decisions, so you can follow the path of any case from open to done.
Changes are deliberate
A workflow only changes when someone asks, and you can review it with your AI agent at any time.
None of this makes an AI agent infallible. That is why the approvals and the record matter: they keep a person in charge of what counts and show what happened in every case.
The test
When to use an agentic workflow
There is a simple test.
Take a piece of work and ask: does it finish the moment it starts, or does it depend on someone coming back to you?
If it finishes straight away and is the same every time, fixed automation is cheaper and faster.
If it depends on a reply, a document, a payment or a decision, and every case is a little different, it is an agentic workflow.
FAQ
Frequently asked questions
What is an agentic workflow?
An agentic workflow is a goal-driven process run by AI: one or more agents plan the steps, use tools to take them, check the results and adapt, with little step-by-step instruction from a person. Agentic business workflows apply this to processes such as chasing invoices or qualifying leads.
What is the difference between agentic and non-agentic workflows?
A non-agentic workflow follows steps fixed in advance, or answers one prompt at a time like a chatbot. An agentic workflow decides its next step itself within your rules, uses tools, and keeps going until the goal is met.
What is agentic workflow execution?
It is running agentic workflows reliably in production: starting them on real events, keeping each case open for as long as it takes, giving the agent tools and instructions, enforcing approvals and recording every step. The software that does this is an execution harness. Askpilot provides one for agentic business workflows.
Do agentic workflows need human approval?
For high-stakes actions, yes. Google describes a person's approval as a mandatory control before actions such as financial transactions or sensitive external messages. In Askpilot you mark those actions in the PROCESS.md, and the agent stops for approval.
What is PROCESS.md?
PROCESS.md is Askpilot's plain-language Markdown format for agentic workflows. One file holds the role, instructions, events, rules, todos and paths, so a person can read it and an agent can execute it. The format is published openly, and Askpilot runs it.
Can I set up agentic workflows from ChatGPT or Claude?
Yes. Connect Askpilot MCP to ChatGPT, Claude or another AI agent and describe the workflow, and your AI agent writes the PROCESS.md and puts it live. You can also create workflows directly in the Askpilot app. Either way, Askpilot's own agent, the Ask agent, runs them.
Do I need an execution harness if my AI agent can run in the background?
Some AI agents can keep working in the background, which suits single tasks. A business process needs more around the agent: triggers from your business tools, cases that stay open for days or weeks, approvals for sensitive actions and a record of every step. That is what an execution harness adds.
Will agentic workflows replace traditional automation?
No. GOV.UK expects traditional and agentic workflows to run side by side. Agentic workflows take the work that varies and waits; fixed automation keeps the identical, high-volume steps.
Invoice past due