Order
The arrows are the order. Branches take a condition; loops run a step over a list.
Agentic workflows
Agentic workflows: fixed steps, judgement inside.
The workflow decides the order. AI reads, judges and drafts inside the steps that need it.
Most business processes are known: the same steps, the same checks, the same hand-offs. What changes is the input. Botminds XFlows keep the steps fixed and auditable, put an agent or a typed decision exactly where judgement is needed, and run them on millions of documents — on schedules, on events, on every new arrival.
- triggerInvoice arrivesa new document in the collection
- codeRead the PDFa Python step
- agentExtract the fieldssupplier, total, due date, lines
- SQLMatch the orderlook up the purchase order
typed decisionPay, hold or reject?
pay0.86
hold for review0.11
reject0.03
below 0.70 confidence → unsure → a person decidesMove to “Approved”then notify the payables team
Move to “Review”a person picks it up in the Inbox
What are agentic workflows?
Agentic workflows are automated processes in which AI agents do the parts that need reading, judgement or writing, inside a sequence of steps the business has defined. The workflow fixes the order, the checks and the hand-offs; the agent handles the messy input. In Botminds an agentic workflow is an XFlow: a graph of steps that run code, call systems, run an agent or make a typed decision — started by a schedule, an event or a new document.
Scales out on KubernetesServices and workers add capacity as work arrives, and give it back when it is done.
Built for millionsMillions of documents, records and agent runs — the indexes and workers grow with them.
Any cloud, any regionAzure, AWS, Google Cloud or on-premises, in the region your data must stay in.
Always onRedundant services, rolling releases and continuous backups, for mission-critical work.
Secure and governedSingle sign-on, encryption, least privilege and a sealed audit trail.
AI workflow vs AI agent
AI workflow vs AI agent: who decides the next step?
In a script, the code decides what happens next. In an autonomous agent, the model does. Agentic workflows sit in between: the business fixes the steps, and AI makes the calls inside them that a rule cannot.
- Scripted automation (RPA)The script decides every step
- Good for
- Same screens, same rules, high volume
- Weak at
- Breaks when a form or a screen changes; cannot read
- Agentic workflowThe workflow decides the order; AI decides inside a step
- Good for
- Known process, messy inputs: read, judge, route
- Weak at
- A path that changes every time
- AI agent with toolsThe model picks the tools and the order
- Good for
- Questions and tasks whose path changes each time
- Weak at
- Less predictable; needs grants, approvals and tests
- Autonomous agentThe model plans, splits and checks a long job
- Good for
- Long jobs a person starts on purpose
- Weak at
- Slowest and costliest per task
Most enterprise work belongs in the second column: an accounts-payable process stays the same from invoice to invoice; the invoices do not. When the path itself changes every time, use an AI agent. When the work is large and spread across teams, run the workflows inside an agent mesh.
Anatomy
Anatomy of an agentic workflow: twenty step types on one canvas.
An XFlow is a graph of steps drawn on a canvas. Each step does one thing; the arrows are the order. Twenty step types on the canvas — AI, code, control, connections and records — plus ready-made intake, classification and industry steps for mesh pipelines.
AIjudgement inside a fixed spine
- Run an agent
- Chunk, embed and index
- Typed decision in a code step
CodePython checked before it runs; commands run with platform secrets kept out
- Python
- SQL
- Shell and JavaScript commands
Controlthe shape of the flow
- Branch on a condition
- Loop over items
- Call another workflow
Connectother systems and flows
- Call an API
- Send an email
- Publish to a topic
- Start on a topic
Recordswrite the result where it belongs
- Move to a stage
- Write table rows
- Fill a report template
- Store a file
- Order a document set
Automationwhat you already have
- Run a Botminds RPA bot
For mesh pipelinesready-made intake, classification and industry steps
- Unpack an archive
- Straighten a scan
- Register a document
- Classify a document
- Finance, trade and lending steps
Input
Each step receives what the step before it returned, or the document the run started on.
Checked early
A step missing a required setting is flagged on the canvas before anything runs.
Logged
Every run keeps each step’s log, so a failure points at the step that stopped the chain.
Workflow orchestration
Workflow orchestration: what starts a workflow, and what every run gives back.
Real workflow orchestration starts on its own — on a clock, a business event or a new document — and every run, however it started, is streamed, logged and controllable.
- A scheduleevery few minutes, hours or days
- A document eventa stage change, new annotations, an export, a re-scored document, a new record
- A new documenta collection’s processing flows run on each one
- A topic messageone flow publishes; every subscriber starts
- A meshas one member of a larger pipeline
- An API call or Piperthe same run surface agents use
- By handa quick run, one document, or custom input
XFlowOne workflowthe same steps, however it started
- Streameach step’s status as it happens
- Pause and resumehold a run, then carry on where it stopped
- Reruna failed run, after the fix
- Cancelstop it, with a confirmation
- Step logsevery step’s lines, level by level
Per document, on arrival
Add your flows after a collection’s processing flow, and they run on each document as it arrives — the most common way agentic workflows go live.
On a business event
A stage change — to the stages you pick — new annotations, an export, a re-scored document or a new record: each can start a workflow.
From your systems
Start a workflow from the API or from a conversation with Piper. Workflows, agents and meshes share one run surface.
Decisions inside a step
Typed decisions: judgement calls, with the code in control.
Many steps need one judgement, not an agent: is this urgent, which document type is it, does this clause allow early termination. A typed decision asks exactly that and returns a typed answer with a probability.
answer = bm.decide(
state={"supplier": s, "total": total, "po_match": ok},
questions={
"action": {
"kind": "choice",
"instructions": "What should happen to this invoice?",
"criteria": ["pay", "hold for review", "reject"],
},
},
purpose="invoice triage",
)
action = answer["answers"].get("action", {})
if not answer["ok"] or action.get("unsure", True):
return {"route": "review"} # a person decides
return {"route": action["choice"]}Pick one
Named options, a probability for each, and a built-in “none of these”.
Place on a scale
Ordered levels: low to critical, no risk to high risk.
Yes or no
The probability that a statement is true.
Unsure is an answer
Below 0.70 confidence the answer is marked unsure, and the code routes it to a person.
Sums stay in code. Dates, totals and contract rules are exact in code. A typed decision is for the call a rule cannot make. Switch the decision model on per workspace.
Agents, workflows, meshes
Agents inside workflows, workflows inside meshes.
The building blocks nest. An agent can be a step of a workflow; a workflow can be a member of a mesh; a workflow can call another. Each keeps its own rules wherever it is placed.
01
An agent as a step
The run-an-agent step calls any agent in the project, with its own tools, model and record. Its answer — typed fields when it has a schema — is the step’s output.
02
A workflow as a member
A mesh runs workflows as members, in thousands of parallel copies. Fan-out, joins and delivery are the mesh’s job.
03
A workflow as a step
One workflow calls another, or publishes to a topic every subscriber starts on. Common steps are built once and reused.
Changing a live workflow with Piper
You
In the invoice workflow, hold anything over 50,000 for review, whatever the decision says.
Agent
I’ll add one code step after Typed decision: if the total is over 50,000, route to Review. It applies from the next run; runs in progress finish as they started. Add it?
You
Yes.
Agent
Added and read back: the step is in place, the code compiles, and the steps after it receive the same input as before.
People in the process
People in agentic workflows: approval at a review stage.
Some decisions must be a person’s. The work is split at a review stage: the first workflow ends by handing the document to a person, and the person’s decision starts the next one — nothing is held open while they decide.
Workflow one
- Agent: draft the decision
- Check the figures
- Move to Review
Review stageA person decidesIt waits in their Inbox — for the personas you pick, with the moves you allow.
ApprovedDeclined
Starts on “Approved”
- Agent: write the letter
- Send it
Starts on “Declined”
- Notify the owner
The stage is the pause; the person’s move is the approval — and it is on the document’s record with who made it and when.
Who may approve
The stage decides: the personas you pick see documents waiting there, with the moves you allow — Approved, Declined.
Where they find it
The document lands in the reviewer’s Inbox, with the agent’s work beside the source.
What is kept
Every move is recorded with who made it and when, next to the runs on either side.
A decision guide
When a fixed workflow beats an agent — and how to combine them.
Five questions, asked in order, settle most designs. The answers usually combine: a workflow with one agent step and one typed decision, split at a review stage, running inside a mesh when volume is high.
1Are the steps known, and the same every time?You could draw the process on a whiteboard today.
YesStart from a workflowgo to question 2
NoUse an AI agentgranted the tools it needs; approvals on anything that changes data
2Does any step need reading or judgement?Unstructured documents, free text, a choice a rule cannot make.
YesAdd AI inside the stepgo to question 3
NoA plain workflowcode, SQL and API steps — exact, cheapest, no model at all
3Is the judgement a pick from known options?Approve or review; which of six document types; urgent or not.
YesA typed decisionone fast call, a probability per option, unsure goes to a person
NoAn agent stepextraction, drafting, research — the agent runs inside the step
4Must a person approve before the next part?A payment, a decision letter, a change to a customer record.
YesSplit at a review stageflow one ends at the stage; the person’s move starts flow two
NoKeep one flowit runs end to end
5Volume, fan-out, or many teams’ steps?Thousands of documents a day; one file into hundreds of parts.
YesRun it in a meshworkflows and agents as members; nothing lost on a crash
NoThe workflow alonetriggered, scheduled or called
Start deterministic
Every step that can be code should be code: exact, cheap and easy to audit.
Narrow the AI
Prefer a typed decision to an agent, and an agent with few tools to one with many.
Measure each AI step
Test each agent step and decision against a golden dataset. See AI agent evaluation.
Agentic workflow examples
Agentic workflow examples, step by step.
Four processes, drawn the way they run. Code steps carry the order; AI steps sit where a document must be read or a call made.
on each new invoice
Invoice to payment
Read the PDF → an agent extracts supplier, totals and lines → SQL matches the purchase order → a typed decision picks pay, hold or reject → hold goes to a person.
- Agent: extraction
- Decision: pay, hold or reject
on each email
Claims intake
Unpack the attachments → classify each file → an agent extracts the claim form → check the policy by API → a typed decision scores urgency → route to the right queue.
- Agent: claim form
- Decision: urgency scale
every Monday
Contract renewal watch
SQL finds contracts ending in 90 days → an agent reads notice periods and auto-renewal → a templated report lists what needs action → email it to the owners.
- Schedule: weekly
- Report: filled template
on a stage change
Supplier onboarding
A supplier moved to Submitted → an agent checks the documents against the checklist → a yes-or-no decision on completeness → complete moves to Review for a person.
- Event: stage change
- Person: final approval
At scale
Agentic workflows at scale: millions of documents, every system feeding in.
Agentic workflows on Botminds are built for internet-scale workloads: events fan in from every connected system, queues absorb the bursts, and workflow runners and workers scale out on Kubernetes — on Azure, AWS, Google Cloud or on-premises, in any region.
Events fan in from every connected system
- SharePoint
- Email inboxes
- Drives
- Cloud storage
- Your API
- Webhooks
- Schedules
- Stage changes
A durable queue absorbs burstsarrivals processed, as workers are added
Workflow runners · scale outStateless orchestration on Kubernetes: any runner picks up any run.
Workers · scale out as load risesThe queue grows, workers are added; it drains, they are given back.
Millions of documents, the same workflows
- stream
- pause
- resume
- rerun
- cancel
- step logs
Every system feeds in
SharePoint, mailboxes, drives, cloud storage, webhooks, your API and schedules all start the same workflows.
Bursts absorbed
A month-end spike lands in a durable queue, and workers are added until it drains.
Runners scale out
Workflow runners are stateless and replicated, so more runs mean more runners, and releases roll through them.
Every run in reach
Pause, resume, rerun or cancel — one run or thousands in bulk — with each step’s log.
- Watch
The Runs tab filters by state, tag and date, with each step’s log.
- Recover
Rerun after a fix, one at a time or in bulk.
- Measure
Compare each change with a pinned baseline on a golden dataset.
- Change
Edit one step in place; it applies from the next run.
- Move
Workflows travel in a solution package to the next environment.
Side by side
Agentic workflows vs RPA vs plain AI agents.
The three are often sold as substitutes. They solve different problems, and most enterprises end up with all three — the question is which one runs each step.
| RPA | Plain AI agent | Botminds agentic workflow | |
|---|---|---|---|
| Who decides the next step | The script | The model | The workflow; AI decides inside a step |
| Unstructured input | Breaks on it | Reads it | Agent and decision steps read it |
| Predictability | High, until a screen changes | Varies run to run | Fixed order; AI where placed |
| Audit | Logs of clicks | A transcript, if kept | Every run, every step, every move by a person |
| A person’s approval | Outside the bot | A prompt in chat | A review stage between flows, with roles |
| Volume | More bot licences | Model calls on every step | Runners and workers scale out; model calls where needed |
| Testing | Re-record the bot | Ad hoc prompts | Golden datasets per agent step and per workflow |
| Your existing bots | — | — | Run a Botminds RPA bot as a step, or call your bot platform through an API step |
Coming from an RPA or workflow platform? See how Botminds compares with UiPath and Automation Anywhere, and the build vs adopt whitepaper.
Questions people ask
Agentic workflows, answered.
More from the platform: the full architecture, and how security works at Botminds.
What are agentic workflows in AI?
Agentic workflows are automated processes where AI agents handle the steps that need reading, judgement or writing, inside an order of steps the business defines. The workflow keeps the process predictable and auditable; the agent handles messy input such as documents and free text.
What is the difference between an AI workflow and an AI agent?
In a workflow, the steps and their order are fixed and AI works inside them. An agent chooses its own next step and tools as it goes. Use a workflow when the process is known and the inputs vary; use an agent when the path itself changes every time.
What is agentic process automation?
Agentic process automation is the step after robotic process automation: instead of scripted clicks, processes run as workflows with AI agents inside them, so they read documents, make bounded decisions and hand exceptions to people.
Are agentic workflows better than RPA?
For processes with documents, emails or judgement calls, yes — RPA breaks when the input changes, while an agent step reads it. For fixed screens and rules RPA still works, and in Botminds you run a Botminds RPA bot as a step, or call your bot platform through an API step.
What are some agentic workflow examples?
Invoice to payment, claims intake, contract renewal watch and supplier onboarding: each has fixed steps, one or two agent steps that read documents, a typed decision, and a review stage where a person approves.
What is workflow orchestration?
Workflow orchestration is starting, ordering, monitoring and recovering the steps of a process. In Botminds a workflow starts on a schedule, a document event, a new document, a topic message, a mesh, the API or by hand, and every run can be paused, resumed, rerun or cancelled.
Can agentic workflows handle millions of documents?
Yes. Events fan in from every connected system into durable queues, workflow runners and workers scale out on Kubernetes, and the same workflows run on every document — on Azure, AWS, Google Cloud or on-premises.
Can a person approve a step in an agentic workflow?
Yes, at a review stage. The first workflow ends by moving the document to review; the person’s move to Approved starts the next workflow, and the move is recorded with who made it.
How do you test an agentic workflow?
Run it against a golden dataset of documents with known answers, repeat runs to check consistency, and compare each change with a pinned baseline. See AI agent evaluation.
When should I use an agent mesh instead of a workflow?
When one item fans out into many parts, or several teams’ workflows and agents must run as one pipeline that survives a crash. The workflows become members of the mesh.
Bring a process. Leave with a workflow.
Show us one process your team runs every day. We’ll draw it as an agentic workflow on your documents — fixed steps, AI where it counts, a person where it must be.