What Are Agentic Workflows and Which 5 Patterns Fit 2026?

An agentic workflow is a system in which a language model and its tools are orchestrated through predefined code paths: engineers decide the sequence of steps in advance, and the model supplies the reasoning inside each step. That single distinction — who chooses the next step, the code or the model — is the whole difference between an agentic workflow and an AI agent, and it is the detail most explanations skip.

TL;DR: if the path is fixed and the model does the thinking at each stop, you have a workflow. If the model decides its own path and keeps deciding until the job is done, you have an agent. Workflows are cheaper, more predictable, and easier to debug; agents are more capable and cost more in latency and money. Most production systems in 2026 are workflows that people call agents.

Key Takeaways

• A workflow orchestrates a model through code paths you wrote; an agent lets the model direct its own process. The line is control over sequencing, not model capability.

• Anthropic documents 5 composable workflow patterns — prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer — which is the most useful taxonomy published for this category.

• Microsoft documents 5 orchestration patterns under different names, and the overlap between the two lists is the strongest signal that these patterns are real engineering rather than vocabulary.

• Every layer of autonomy you add buys capability with latency and cost; the published guidance from both vendors is to use the least complex option that reliably works.

• Decision rule: if a single well-prompted model call with retrieval solves your task, you do not have an agentic workflow problem — you have a prompt to write.

• Decision rule: reach for multiple agents only when a single agent genuinely cannot hold the task, and cap collaborative patterns at 3 or fewer participants.

Workflow or Agent? The Line That Actually Matters

The cleanest statement of the distinction comes from the engineering team at Anthropic, and it is worth reading twice because both halves matter. In a workflow, "LLMs and tools are orchestrated through predefined code paths." In an agent, systems are those "where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

Both sit on a shared foundation. Anthropic's engineering guide describes the basic building block of agentic systems as a language model enhanced with augmentations such as retrieval, tools, and memory — the "augmented LLM." Everything above that block is a choice about who holds the steering wheel.

Other vendors define the autonomous end of that spectrum in compatible terms. According to AWS, agentic AI is "an autonomous AI system that can act independently to achieve pre-determined goals," operating across 4 stages it names as perceive, reason, act, and learn. Notice that AWS's definition assumes the goal is given and the path is not — which is exactly the agent side of Anthropic's line.

The Five Composable Patterns, in Plain English

The reason this taxonomy is worth memorising is that almost every real system is one of these five, or two of them stacked. They are composable by design, and none of them requires a model to choose its own path.

That set is not our invention and should not be quoted as such. According to Anthropic's engineering guide, those 5 patterns plus the augmented-LLM building block cover the systems the team has seen work in production, and the article is explicit that the patterns are meant to be combined rather than chosen exclusively.

Anthropic documents 5 composable workflow patterns, plus 1 building block. Prompt chaining and routing between them account for a large share of what teams actually ship, because both are fully deterministic in their sequencing — the model never chooses where the data goes next, which makes both patterns straightforward to test.

A Second Vocabulary: How Microsoft Names the Same Ideas

If you read two vendors' documentation in the same week you will meet the same patterns under different names, and the aliases are worth knowing so you do not think you are looking at two different fields. According to Microsoft's Azure Architecture Center, 5 orchestration patterns cover multi-agent architectures: sequential, concurrent, group chat, handoff, and magentic.

Microsoft's guidance on the collaborative pattern carries a number worth remembering. According to Microsoft, managing conversation flow and preventing infinite loops gets harder as agents are added, and teams should consider limiting group chat orchestration to 3 or fewer agents.

Cap collaborative patterns at 3 or fewer agents. The documentation also names a specific and very useful sub-pattern: the maker-checker loop, where one agent proposes and another evaluates against defined criteria, pushing work back with feedback until the checker approves or an iteration limit is hit.

That iteration limit is not a detail — it is the thing standing between your workflow and an unbounded bill. US Tech Automations sets an explicit cap on every critique loop it runs, so a drafting step that keeps failing its checker halts and raises a flag instead of quietly re-billing the same page.

Vocabulary Check: Eight Terms You Will Hit in Vendor Docs

Straight-through processing is the metric that ends most arguments about whether a workflow is working. It is not a model quality score — it is the percentage of real items that finished without a person, which is the number a finance or operations lead will actually ask you for.

One Pattern, Walked End to End

Here is prompt chaining and routing stacked, on a real payment event. A subscription business processing 1,240 successful payments a week wires a workflow to Stripe's payment_intent.succeeded event: the event fires, step 1 pulls the customer's plan and support history, step 2 classifies the payment as a first charge, a renewal, or a recovered failure, and step 3 drafts the matching message — a 3-branch route where roughly 60 first-charge events a week get an onboarding sequence and the rest get a receipt.

According to Stripe's API event reference, that event "Occurs when a PaymentIntent has successfully completed payment" — a reliable trigger rather than a polled guess, and 1 of the named platform events these workflows hang from.

Nothing in that description required the model to decide its own path. The branches were enumerated by an engineer, the classification step was the only place judgement entered, and the whole thing can be replayed against last week's events to check it still behaves. That is what makes it a workflow, and why it is testable in a way an open-ended agent is not.

This is the level US Tech Automations builds at: a named platform event as the trigger, a fixed set of branches, a model doing classification and drafting inside those branches, and a human approval step before anything customer-facing leaves the building. Our own agentic workflows platform exists to run exactly that shape — orchestration you can enumerate, not autonomy you have to trust.

For the surrounding decision, agentic agents explained covers what changes operationally once these patterns are in production.

Choosing the Least Complex Thing That Works

Both vendors converge on the same warning, which is the strongest signal in this whole article. Microsoft frames it as a ladder: use the lowest level of complexity that reliably meets your requirements, where the levels run from a direct model call, to a single agent with tools, to multi-agent orchestration.

Anthropic's version of the same advice is blunter about the cost. According to Anthropic, agentic systems "often trade latency and cost for better task performance," and for many applications optimizing single LLM calls with retrieval and in-context examples is usually enough. Read that as permission: choosing not to build an agentic workflow is a legitimate and often correct engineering decision.

Governance sets a floor under the same choice. According to NIST, the AI Risk Management Framework released on January 26, 2023 organises the work into 4 functions — Govern, Map, Measure, and Manage — and a system whose path you cannot enumerate is materially harder to measure and manage than one whose path you can.

Nobody Is Answering This Query Yet

One reason this page exists is that the demand is measurable and the supply is not. Across our own Search Console property between January 1 and May 20, 2026, 40 distinct queries containing "agentic" produced 1,156 impressions and zero clicks, with the head term "agentic workflows" alone accounting for 651 of those impressions at an average position of 52.6.

Our own data: 1,156 impressions, 0 clicks, across 40 agentic queries. An average position in the fifties means the pages exist somewhere in the index and are being shown, not read — which for a definitional query usually means the definition is buried under a product pitch.

US Tech Automations publishes under the same constraint it is describing. Every post clears a fixed set of automated quality checks before publication, and a re-count of the published library on July 28, 2026 returned 14,315 blog markdown files on the live branch — a corpus assembled by the workflow patterns described above rather than by a single model call.

Who Should Care, and Who Should Wait

This page is for engineering and operations leads at firms somewhere past their first LLM feature — a team that already has a working model call in production, is now looking at a multi-step process such as document intake, support triage, or content review, and needs to decide whether that process wants a workflow, an agent, or neither. If you can name the steps of your process on a whiteboard, you are ready to pick a pattern.

Red flags: Skip this decision entirely if your task is a single-step classification or summarisation that a well-written prompt already handles, if you have no evaluation criteria yet for what "correct output" means, or if no one owns the error cases — in all three situations adding orchestration buys complexity you cannot yet measure.

Once the pattern choice is made, business workflow automation compared and the step-by-step build guide cover the implementation side.

Frequently Asked Questions

What is an agentic workflow in simple terms?

It is a multi-step process where the steps are fixed by code and the thinking inside each step is done by a language model. A document intake pipeline that reads a file, classifies it, extracts fields, and routes exceptions to a person is an agentic workflow: a human designed that sequence, and the model supplies judgement at each stage rather than deciding what happens next.

What is the difference between an agentic workflow and an AI agent?

Control over sequencing. Anthropic's distinction is that a workflow orchestrates models and tools "through predefined code paths," while an agent is a system where models "dynamically direct their own processes and tool usage." A workflow can be replayed step by step; an agent may take a different route on every run, which is both its advantage and the reason it is harder to test and budget for.

Are agentic workflows and agentic AI the same thing?

No, though the terms are used interchangeably in marketing. "Agentic AI" typically describes the autonomous end of the spectrum — AWS defines it as an autonomous system acting independently toward pre-determined goals across perceive, reason, act, and learn stages. "Agentic workflow" describes the orchestrated end, where the path is fixed. Most products sold as agentic AI are, structurally, agentic workflows.

How many agentic workflow patterns are there?

Anthropic documents five composable patterns — prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer — plus the augmented-LLM building block underneath them. Microsoft documents five orchestration patterns for multi-agent systems using different names. The two lists overlap heavily, so learning either one is enough to read the other vendor's documentation without getting lost.

When should I not use an agentic workflow?

When a single model call already works. Anthropic's guidance is explicit that agentic systems trade latency and cost for task performance, and that optimizing single LLM calls with retrieval and in-context examples is usually enough for many applications. If your task is one step, has no branching, and passes evaluation with a good prompt, orchestration adds cost and failure modes without adding capability.

Do agentic workflows need human approval steps?

Not technically, but most production systems that touch customers or money include one. A human-in-the-loop step is where a person approves, corrects, or rejects the model's output before it takes effect, and it is the standard control for anything irreversible — an outbound email, a ledger entry, a refund. Adding one costs latency and buys you a bounded failure mode.

What is a maker-checker loop?

A specific collaborative pattern where one agent produces work and a second evaluates it against defined criteria, sending it back with feedback until the checker approves or a maximum iteration count is reached. Microsoft documents it as a form of group chat orchestration and recommends capping collaborative patterns at three or fewer agents, since loop control gets harder with each participant added.

Sources and Further Reading

The primary sources for this page, each read directly on July 28, 2026:

Deliberately absent from this page: any market-size figure, adoption percentage, or productivity statistic for agentic AI. Several are widely quoted, none could be verified against a primary source today, and a definition page that leans on an unverifiable number to sound authoritative undermines the definition it came to give.

If you have a named platform event and a process you can draw on a whiteboard, you have the two ingredients an agentic workflow needs — the pattern catalogue above is how you pick the shape. The patterns US Tech Automations runs in production, and what each is suited to, are described at ustechautomations.com/platform/agentic-workflows.