the mechanics guide

Agentic workflows, explained

An agentic workflow is a defined process in which an AI agent does the judgment work inside each step. This guide explains the two architectures behind the phrase, pipeline and conversation-first, and how steps, gates, decisions, loops and child tasks fit together into a process an AI can run again and again. In ConvOps, that process is a workflow stored in your workspace, which the agent reads one step at a time.

the route

10 sections · 12 min

  1. Two architectures
  2. Steps
  3. Gates
  4. Decisions and detours
  5. Loops
  6. Fan out, then wait
  7. Five patterns
  8. Versus automation
  9. Across teams
  10. Choosing
01Two architectures

What are the two kinds of agentic workflow?

When people say "agentic workflow," they mean one of two designs. The first is the pipeline, the model of frameworks such as LangGraph: a graph written in code where each node invokes an agent, output flows to the next node, and the framework owns control flow. The second is conversation-first: the process is stored as structure, the agent works in its own environment (a coding client, a chat), and an engine feeds it one step at a time and decides whether it may advance.

Pipelines shine when the task is fixed, high-volume, and headless: classify, extract, transform. The conversation-first style shines when the work is the kind humans do in sessions: build a feature, produce an article, run a sales motion, because the human can be in the loop natively, the agent keeps its full working context, and the process survives across sessions and executors.

Pipelinecode owns the flow
Conversation-firstan engine hands out steps
fig 1 Same phrase. Who owns the flow is the difference.
02Steps

What is a step in an agentic workflow?

The atom of a workflow is the step: a named unit with instructions for exactly that unit. The delivery discipline is what separates a step from a prompt section: the agent receives the current step when the work reaches it, not the entire process up front. A model holding one instruction executes better than a model juggling forty, and an operator editing one step fixes every future run.

Steps are also where organizational knowledge stops being tribal. "Reproduce before fixing" or "cite every claim" written into a step is process that executes, not advice that depends on who is driving.

reproducestep 1
fixstep 2
teststep 3
done
fig 2 The AI holds one step at a time, not the whole manual.
03Gates

What is a gate in an agentic workflow?

Between steps sit the controls. A gate is a condition evaluated by the engine on every attempted advance: approval given, children complete, status reached, context recorded. The run does not move until the condition is true, and a refused advance says exactly what is missing.

This is the mechanism that lets a workflow make promises: "nothing ships without sign-off" is enforceable when sign-off is a gate, and merely aspirational when it is a sentence.

in the prompt read by the AI

"wait for approval"

Shipped. Nobody said yes.

as a gate checked by the engine

Waiting for approval.

fig 3 A sentence asks. A gate holds.
04Decisions and detours

How do agentic workflows branch?

Real processes branch: a bug that cannot be reproduced closes instead of proceeding; unclear scope routes to research first. A robust way to model this is decisions that evaluate real fields of the work and, when a condition matches, detour into a sub-workflow that runs to completion and returns.

The detour design keeps something valuable: the main line of the run stays single and auditable. You can always answer "where is this run?" with one cursor, even while a branch executes its own steps and gates one level down.

triage
scope unclear?
researchsub-workflow
back on line
done
fig 4 A matched condition takes a detour that always comes back to the line.
05Loops

How do loops get work to done?

The most important control flow in quality-bearing work is the fix loop: review finds problems, work returns to the fix step, and the cycle repeats until the review passes. In a conversation-first engine, this is the return trip from an invoked sub-workflow pointing backward: findings send the run back, the gate ahead only opens on green.

Loops are also where the honesty of a workflow shows. A pipeline that cannot loop ships whatever the last node produced. A workflow that loops until a gate passes has a definition of done that is enforced, not hoped for.

findings: back to fix
fixreviewsign-offdone
fig 5 Review sends it back until the gate opens on green.
06Fan out, then wait

How do agentic workflows run work in parallel?

Some pieces of work do not depend on each other: the fix, its tests, its docs. There are two ways to run them side by side. A step's instructions can tell the orchestrator, your AI session, to start several agents at once: the engine holds the step, the session does the fan-out. Or the work splits into child tasks: children that share an order start together, each with its own agent, potentially on its own runner, and the parent's wait_for_children gate holds until all are done. The process stays one process; the execution fans out and comes back.

Decoupling the process from its executors is what makes the whole design durable: the same workflow can run attended in a person's coding client today and unattended on a scheduled backend next month, because the engine owns the steps and the executors own only the compute.

Child tasksfix, tests, docs start together
Own runnerseach child can run elsewhere
Parent waitsuntil all children are done
fig 6 The process stays one process. The execution fans out, the parent waits.
07Five patterns

What are the common agentic workflow patterns?

Most agentic workflows are built from a handful of patterns. Naming them helps when you design a new process, because each one answers a different question: how is the work planned, how is it checked, how is it routed, how is it split, and where does a person decide. Real workflows combine several, and the steps, gates, decisions, loops and child tasks above are the parts each pattern is made of.

The table pairs each pattern with an everyday example and the part of a conversation-first engine that carries it. A pattern does not need its own framework. It needs a place in the process where the engine, not the model, decides what happens next.

Five agentic workflow patterns, with the part of the engine that carries each.
PatternWhat it doesExampleCarried by
Plan and executeOne step writes the plan, later steps carry it outA research brief written first, then each section draftedOrdered steps, with the plan recorded as context
Reflection or fix loopA review sends the work back until it passesAn article returned to the writer until the editor's checks passA fix loop with a gate ahead of it
RouterA decision picks the path from the facts of the workA support ticket routed to refunds, bugs or how-to answersA decision that starts a sub-workflow and returns
Orchestrator and workersOne parent splits the work and waits for the partsA quarterly report split into one task per regionChild tasks and a wait_for_children gate
Approval gateThe step holds until an explicit approval is givenA price change held for the sales leadA requires_approval gate on that step
08Versus automation

How is an agentic workflow different from automation or a single agent?

Three things get called agentic, and they solve different problems. Classic automation, the kind built in tools such as Zapier or n8n, moves data between apps along fixed paths: when a form arrives, create a record, send a message. A single agent is a model with tools and a goal, working out its own steps as it goes. An agentic workflow sits between them: the steps are fixed and written down, and the judgment inside each step is the agent's.

The choice follows the work. If no step needs judgment, automation is cheaper and more predictable. If the work is one-off exploration, a single agent with good tools is enough. If the work repeats, needs judgment and has a quality bar or a person who must decide, an agentic workflow is what keeps run 30 as good as run 3.

Automation, a single agent and an agentic workflow, compared.
QuestionClassic automationA single agentAn agentic workflow
Who decides the next step?The fixed pathThe modelThe stored process
Where is judgment used?NowhereEverywhere, uncheckedInside each step, checked between steps
How does it stop for a person?A separate approval appIf the model remembers to askA gate the engine checks
What happens when it is wrong?It repeats the error until someone edits the pathEach run fails differentlyA fix loop sends the work back, and one step edit fixes future runs
What does it leave behind?Run logsA transcriptA cursor, the step history and the recorded metrics
09Across teams

What do agentic workflows look like across teams?

The same parts show up in very different work. The examples below are typical shapes, each labelled with the kind of team that would run it. In each one, the agent does the reading, drafting and checking, and the workflow decides the order, the checks and the stops.

In ConvOps, any of these can be written as a workflow, or started from a template, by asking your AI client over MCP. The agent sends email, updates a CRM or publishes a page only through the tools its environment gives it, so the workflow controls when those actions happen and the environment controls whether they can happen at all.

  1. Enterprise, regulated content: a medical claims check is one fragment embedded in every content workflow, and a legal reviewer approves before anything is published.
  2. Finance operations: a month-end reconciliation matches entries, records the count of unmatched items with the export as evidence, and holds for a controller before adjustments are posted.
  3. Sales: an account research workflow builds a brief, drafts a first message, and holds at a gate until the account owner approves the draft.
  4. Support: a triage step routes each ticket; refund requests detour into a sub-workflow with an approval, and the rest return to the main line with a drafted reply.
  5. Research: a literature review splits into one child task per source set, and the parent waits until every summary is in before writing the synthesis.
  6. Software delivery: a bug fix reproduces the bug, fixes it, and loops through review until the tests and the reviewer pass, then waits at the merge.
10Choosing

Which architecture should you choose?

A short decision aid:

  1. Fixed transformation at volume, no humans mid-flow: build a pipeline
  2. Session-shaped work, humans sometimes in the loop, quality bar enforced by review: conversation-first engine
  3. Need the same process attended today and unattended later: conversation-first, since the process must outlive the driver
  4. Need per-run auditability with one answer to "where is this run?": an engine that keeps a single cursor

Frequently asked questions

What is an agentic workflow?

An agentic workflow is a defined process that an AI agent runs step by step, doing the judgment work inside each step while an engine decides whether the run may move on. In ConvOps, the workflow is stored in your workspace, the agent receives only the current step with its rules, and gates, decisions and loops control what happens between steps.

What is the best agentic workflow pattern?

There is no single best pattern, and ConvOps lets one workflow combine all five, so each is used where it fits. Plan and execute suits long tasks, a fix loop suits work with a quality bar, a router suits mixed incoming work, orchestrator and workers suits work that splits, and an approval gate suits expensive or irreversible steps.

How is an agentic workflow different from Zapier or n8n?

Zapier and n8n move data between apps along a fixed path, while an agentic workflow gives the judgment inside each step to an AI agent. ConvOps keeps the order, the checks and the stops in the process, and the agent works in the AI client you already use. The engine checks each advance before the run moves on.

Can an agentic workflow run with nobody watching?

Yes, if the process holds the stops. ConvOps runs the same workflow attended in your client or unattended on an executor started by a schedule, and a run waits at a gate until an explicit approval is given while other work keeps moving. A run with nobody at the keyboard cannot ask in a chat, so the checkpoints must be part of the workflow.

How do I build an agentic workflow in ConvOps?

Connect your AI client to ConvOps over MCP, then describe the process and ask the client to create the workflow, or start from a template. Name each step and its instructions, add gates where a person, a metric or a report should decide, and add decisions or loops where the work branches. Then create a task on it and run the first step.

Who should use agentic workflows?

Teams that already hand real, repeatable work to AI agents, such as bug fixes, content reviews or a month-end close, and want the same steps every time. ConvOps is built for them: the process lives in a workflow template, the agent receives one step at a time, and gates hold the points where a person, a metric or a report should decide.