Why Agentic Workflows Make AI More Useful: Process, Parallelism and Modularity

On this page

1. 1. Performance: Better Results Come From Better Process
2. Why Reviewing Works
3. Smaller Models Can Be Enough
4. 2. Parallelism: Doing Several Parts of the Job at Once
5. Why Not Use One Large Agent?
6. Where Parallel Work Can Go Wrong
7. 3. Modularity: Replace One Part Without Rebuilding Everything
8. A Simple Business Example
9. The Role of Standards
10. Adoption Is Moving Faster Than Implementation
11. Agents Beyond Text
12. Making Agentic Systems More Reliable
13. When Agentic Workflows Are Not Worth It
14. Glossary
15. The Bottom Line
16. References
17. Footnotes

Most people use AI in a simple way: they type a question, wait a few seconds, and get an answer.

That is direct prompting. It can be useful, but it has a built-in limit. The model has one chance to answer, with whatever information and reasoning it can produce in that moment.

An agentic workflow gives the model a process instead of asking for a single answer. The system can plan, search, use tools, review its own work, correct mistakes, and try again. In some cases, it can divide the work among several agents that operate at the same time.

A simple way to think about it is this: direct prompting is asking someone to answer from memory. An agentic workflow is asking that person to research, draft, check, revise, and then submit the answer.

That difference matters because many valuable tasks are not one-step tasks. Writing reliable code, researching a complex topic, comparing sources, preparing a report, processing a claim, or checking a compliance issue all require several steps. Agentic workflows are useful because they are designed around that reality.

They improve AI systems in three main ways: better performance, faster execution through parallel work, and easier improvement through modular design.

Process beats a bigger model

1. Performance: Better Results Come From Better Process#

One of the clearest examples comes from coding.

A common benchmark for testing AI coding ability is HumanEval. It gives a model small programming tasks and checks whether the code works. For example, the model might be asked to write a function that finds the median of a list, checks whether a number is prime, or removes duplicates from a sequence.

The score often used is called pass@1. In plain language, that means: how often does the model get the answer right on its first try?

Andrew Ng’s team at DeepLearning.AI collected the figures that made this argument popular. Asked to solve these coding tasks directly, GPT-3.5 scored 48.1% and GPT-4 scored 67.0%. That made sense. GPT-4 was a stronger model. But GPT-3.5 placed inside an agentic workflow, told to write the code, review it, look for errors, and improve it, reached as high as 95.1%.¹

Two caveats are worth stating plainly, because the number gets quoted without them. Those figures were gathered from several published agent methods rather than produced by one controlled head-to-head run, and the 95.1% comes from combining agent techniques, not from reflection alone.

Even with those caveats, the shape of the result holds: an older model, with a better process around it, could outperform a newer model answering directly.

The point is not that GPT-3.5 is better than GPT-4. It never was. What the result shows is that a structured process can add a lot of capability. In some cases the process matters as much as the model, and sometimes more.

Consider a simple coding example. Suppose the task is to write a function that finds the median of a list of numbers.

A direct prompt might produce code that sorts the list and returns the middle value. That works when the list has an odd number of items. But it fails when the list has an even number of items, because the correct median should be the average of the two middle values.

An agentic workflow can catch that. The first step writes the code. The second step asks the model to review the code and identify inputs that could break it. The model may then notice the even-length case. A third step rewrites the function to handle both odd and even lists.

The model did not become more intelligent between step one and step three. It was given a chance to inspect its own work from a different angle.

This is close to how people actually work. Most people do not write a difficult memo, contract, spreadsheet formula, or piece of code perfectly on the first attempt. They create a version, test it, find problems, and revise. Agentic workflows bring that same pattern to AI systems.

Two different jobs

Why Reviewing Works#

It may sound strange that a model can find a mistake in its own answer. If it knew the answer was wrong, why did it make the mistake in the first place?

The reason is that producing an answer and judging an answer are different tasks.

When a model generates text or code, it is building forward. Each word or line influences the next one. Once it starts down a path, it may continue even if the path is flawed. When it reviews an answer, the task is narrower. It can look at the whole output and ask: does this satisfy the requirement?

People behave the same way. It is easier to spot a mistake in a finished paragraph than to write the paragraph perfectly on the first try. It is also easier to debug code after seeing the complete version than to avoid every bug while typing.

Agentic workflows use that difference. They split work into stages: create, inspect, improve.

Not every step needs the big model

Smaller Models Can Be Enough#

A second lesson has become more important as companies try to put agents into production: not every step needs the most powerful model available.

Many agentic tasks are constrained. That means the model is not being asked to write a novel or solve an open-ended philosophical problem. It may be asked to choose from a short list of tools, fill in a fixed data structure, summarize one page, classify a request, or extract three facts.

Two terms matter here.

A schema is a fixed format for information. For example:

name: Sarah Lee

invoice_amount: 840

payment_due_date: 12 August 2026

The model does not have to invent a format. It has to put the right information into the right fields.

An API is a defined way for software systems to talk to each other. In an agentic workflow, an API might give the model a few permitted actions, such as:

search_database

send_email

create_ticket

check_order_status

run_code

That is very different from asking the model to do anything at all. The model has a menu of allowed actions.

Because many workflow steps are narrow like this, smaller models can often handle them well. A company might use a strong model for planning, then use smaller, cheaper models for routine extraction, formatting, classification, or checking.

That can matter financially. If one user request triggers forty model calls behind the scenes, using an expensive model for all forty calls may be wasteful. A more practical design uses the strongest model only where it changes the result.

A typical setup might look like this:

• A strong reasoning model creates the plan.
• A smaller model reads documents and extracts facts.
• A search tool retrieves current information.
• A code tool runs calculations.
• A writing model prepares the final response.
• A checking step verifies whether the answer follows the original request.

This is where the real advantage appears. The system is no longer just a chatbot. It becomes a process made of specialized parts.

Doing several things at once

2. Parallelism: Doing Several Parts of the Job at Once#

The second benefit is speed.

Parallelism means doing multiple things at the same time. Sequential work means doing one thing after another.

Humans are mostly sequential when doing research. A person searches, opens a result, reads it, goes back, opens another result, reads that, and continues. Even a fast reader can only process one page at a time.

Software can do something different. It can send several searches at once, open many pages at once, summarize them at once, and then combine the results.

Take a simple example: writing an article about recent developments in black hole physics.

A human researcher might start with one search query, read a few results, think of a better query, read more results, save useful links, compare sources, and then begin writing. That could take an hour or more.

An agentic workflow can divide the work.

First, three agents generate search queries from different angles. One looks for recent telescope findings. Another looks for theoretical work on black holes and quantum physics. A third looks for news about gravitational waves or primordial black holes.

Then the system runs those searches at the same time.

If each search returns three useful pages, the workflow now has nine pages to inspect. A person would read those one by one. The workflow can fetch all nine at once and assign each page to a separate summarizing step.

Each summarizer extracts the useful claims, dates, names, and uncertainties. The final step receives nine short summaries rather than nine full web pages. It then writes a draft that draws from all of them.

Interactive // nine sources, two ways

Same nine sources, same nine reads, same closing synthesis. The only difference is whether they wait for each other. Advance the clock and watch where the two lanes part.

Next tickRun to the endReset

Shared clocktick 0

One at a time

• Telescope resultsreading
• Quantum gravity paperwaiting
• Gravitational waveswaiting
• Primordial black holeswaiting
• Observatory releasewaiting
• Conference abstractwaiting
• Textbook chapterwaiting
• Science explainerwaiting
• Preprint listingwaiting

synthesisewaiting

elapsed0 ticks

model calls0

Reading source 1 of 9. The other eight are still shut.

All at once

• Telescope resultsreading
• Quantum gravity paperreading
• Gravitational wavesreading
• Primordial black holesreading
• Observatory releasereading
• Conference abstractreading
• Textbook chapterreading
• Science explainerreading
• Preprint listingreading

synthesisewaiting

elapsed0 ticks

model calls9

All nine sources opened in the same tick.

Both lanes end on 10 model calls. Running them at the same time saves clock, not work, and it does not make any of it free.

This is slower than a single chatbot response. It might take two minutes instead of five seconds. But the five-second answer is not doing the same job. It is answering from memory or from limited context. The workflow is researching, filtering, and synthesizing.

The fair comparison is not between a five-second chatbot answer and a two-minute workflow. It is between a two-minute workflow and a human doing the same research properly.

Why the work gets split up

Why Not Use One Large Agent?#

There is a practical reason to split the work: models have limited working memory.

The technical term is context window. It means the amount of text the model can consider at once, including the user’s prompt, previous messages, documents, tool results, and the model’s own output.

If you feed a model nine long articles at the same time, several problems appear. It becomes expensive. It becomes slower. Important details may be buried in the middle of the input. That last one is not a hunch: researchers testing how models use long contexts found performance is highest when the relevant information sits near the beginning or the end, and degrades significantly when the model has to reach into the middle.² Even when a model can technically accept a long input, it may not use every part of that input equally well.

A better approach is to give each agent a smaller job.

One agent reads one document. Another reads another document. Each returns a clean summary. The final model works with those summaries.

This reduces clutter. It also makes the system easier to debug. If the final answer includes a wrong claim, you can often trace it back to the source summary that introduced the error.

The honest half

Where Parallel Work Can Go Wrong#

Parallelism creates new problems.

Different agents may return conflicting information. One source may say a finding is confirmed, while another says it is preliminary. The final agent has to notice the conflict and handle it carefully.

The system may also combine good summaries into a weak final answer. Synthesis is its own skill. A pile of accurate notes does not automatically become a strong article.

There is also a cost issue. Ten model calls cost more than one model call. Running them at the same time saves clock time, but it does not make them free.

The coordinator matters too. In many multi-agent systems, one agent decides what work should be assigned to the others. If that coordinator misunderstands the task, the rest of the system may execute the wrong plan efficiently.

For this reason, strong agentic systems usually include checkpoints. They ask whether the plan matches the user request, whether sources conflict, whether required information is missing, and whether the final output actually answers the question.

How much freedom the system gets in the first place is its own decision, and a large one. AI Agents Are a Spectrum, Not a Switch covers that axis: the more latitude an agent has to choose its own next step, the more control the design around it has to supply.

Replaceable parts

3. Modularity: Replace One Part Without Rebuilding Everything#

The third advantage is modularity.

A modular system is built from parts that can be changed independently. In an agentic workflow, search, planning, summarization, writing, code execution, database access, and final review can all be separate components.

That matters because the AI ecosystem changes quickly. New models appear. Search providers improve. Tool protocols become more standardized. Prices change. If a workflow is modular, a team can replace one component without rebuilding the entire system.

Consider the research example again.

One step is web search. That step could use Google through an API provider, Bing, DuckDuckGo, Brave Search, Exa, Tavily, You.com, or a specialist academic search tool. These tools do not return the same results. Some are better for current news. Some are better for technical papers. Some are cheaper. Some return cleaner text. Some are better at finding obscure sources.

If the task is to explain black holes to general readers, a general search engine may be enough. If the task is to summarize new papers published this month, a scholarly or news-oriented search tool may be better.

A modular workflow can swap the search component while leaving the planning, summarizing, and writing steps unchanged.

The same applies to models.

A model that is excellent at planning may not be the best model for writing polished prose. A model that writes well may not be the cheapest or fastest option for extracting dates and figures from a document. A code-focused model may be better than a general model for debugging.

In practice, a production workflow might use different models for different jobs:

• One model to plan the task.
• Another model to search and extract facts.
• Another model to write the final answer.
• Another model to check whether the answer follows policy or formatting rules.
• A code interpreter to run calculations instead of asking a language model to do arithmetic from memory.

This makes the system cheaper, faster, and easier to improve. If a better extraction model appears, you replace that step. If a better writing model appears, you replace the writing step. The rest of the workflow stays intact.

What it looks like at work

A Simple Business Example#

Imagine a company wants to automate customer refund requests.

A direct chatbot might read the customer’s message and reply with a guess: approved, denied, or please wait.

An agentic workflow can do more.

It can classify the request, look up the customer’s order, check whether the item was delivered, read the refund policy, inspect whether the customer is inside the return window, check for fraud signals, draft a response, and escalate uncertain cases to a human.

Several of those steps can happen in parallel. The order lookup, fraud check, policy retrieval, and customer history check can run at the same time.

The system is also modular. If the company changes payment providers, it replaces the payment lookup tool. If the refund policy changes, it updates the policy source. If a better fraud model becomes available, it swaps that component.

That is the practical value of agentic design. The gain is not that the model sounds smarter. The gain is that AI is connected to the actual steps required to complete the work.

Shared interfaces

The Role of Standards#

For modular systems to work well, tools need standard ways to connect.

One important development is the Model Context Protocol, often called MCP, introduced and open-sourced by Anthropic in November 2024.³ It standardizes how AI systems connect to tools, databases, files, and services. Before protocols like this, every connection required custom integration work. Connecting an AI system to a calendar, database, code repository, or messaging tool meant writing separate glue code each time.

MCP aims to make those connections more reusable. If a tool supports the protocol and an agent system supports the protocol, they can connect with less custom work.

Another development is Agent2Agent, often called A2A, launched by Google in April 2025 and donated to the Linux Foundation that June, where it now runs as a vendor-neutral project with AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow among its founding members.⁴ Its goal is to help agents communicate with other agents across platforms. That matters because a future workflow may not use only one agent from one provider. It may involve several agents, each built for a different task.

The broader pattern is clear: agentic AI becomes more useful when models, tools, and agents can connect through shared interfaces.

The gap between interest and delivery

Adoption Is Moving Faster Than Implementation#

Large companies are actively testing agentic systems. Gartner expects 33% of enterprise software applications to include agentic AI by 2028, up from less than 1% in 2024.⁵ Many software vendors are already adding agents to products for customer service, sales, coding, analytics, finance, HR, and operations.

The interest is easy to understand. Many business processes consist of repeated steps: read a request, find the relevant data, apply a rule, update a system, write a response, and record what happened. Those are natural targets for agentic workflows.

The harder part is implementation, and the same analyst has a number for that too. Gartner also predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, on cost, unclear business value or inadequate risk controls.⁵

Many companies have old systems that were never designed for AI agents. The data may be scattered across PDFs, emails, spreadsheets, internal portals, and legacy databases. Some systems lack modern APIs. Others require human credentials. Some cannot be updated in real time. In many organizations, even finding the right document is difficult.

An agent cannot reliably act on information it cannot access. It also should not be given broad permissions without controls. A serious deployment needs scoped access, logging, approval steps, and a way to undo or review actions.

Data quality is another obstacle. If product data is inconsistent, customer records are duplicated, policies are outdated, or documents contradict one another, the agent will inherit those problems. Better prompting will not fix a broken information environment.

This is why the most successful deployments usually involve redesigning the workflow, not just adding a chatbot to the old one. The process has to be shaped around what agents can do well: retrieve information, call tools, compare options, draft outputs, check rules, and escalate uncertain cases.

Out of the chat window

Agents Beyond Text#

Agentic workflows are not limited to writing documents or answering questions. They are beginning to appear in scientific research and physical automation.

In chemistry, for example, researchers have connected language models to tools that search reaction databases, propose synthetic routes, schedule equipment, and interpret experimental results. One such system, Coscientist, was given internet and documentation search, code execution and experimental automation, and used them to design, plan and run experiments, including optimising palladium-catalysed cross-coupling reactions.⁶ A system like that can plan a reaction, send instructions to lab equipment, read measurements, and adjust the next attempt.

Some of the equipment involved needs translation for non-specialists.

A spectrometer helps identify what a substance is made of. A liquid-handling robot moves precise amounts of liquid from one container to another. Chromatography equipment separates mixtures so researchers can analyze their components.

The agent does not replace the laws of chemistry. It coordinates the steps: search, plan, run, measure, revise.

Drug discovery is another area where AI-driven workflows are gaining attention. Insilico Medicine’s ISM001-055, now named rentosertib, is an anti-fibrotic candidate for idiopathic pulmonary fibrosis whose target and molecule were both identified using generative AI. Phase IIa results were published in Nature Medicine in June 2025: a 12-week trial in 71 patients across 22 sites, where the 60 mg once-daily arm showed a mean improvement in lung function of 98.4 mL against a decline of 20.3 mL on placebo.⁷ A Phase III trial has since begun.⁸ It is a candidate in trials, not an approved medicine, and the trial was small.

This kind of example shows where agentic workflows may become most valuable. They can connect reasoning, search, specialized tools, and repeated experimentation.

Sources and a paper trail

Making Agentic Systems More Reliable#

Two ideas are especially important for reliability: grounding and auditability.

Grounding means tying the model’s answer to real sources or tool results. The most common method is retrieval augmented generation, usually shortened to RAG.⁹

RAG works by retrieving relevant documents before the model answers. Instead of asking the model, “What is our refund policy?”, the system first finds the refund policy document and then asks the model to answer from that document. If you want the mechanism in full, that is Retrieval Augmented Generation, and What RAG Is Actually Used For covers where it earns its keep.

This reduces hallucination. A hallucination is when a model produces a confident answer that is not true. The answer may sound polished, but it is unsupported or fabricated.

RAG does not make hallucinations impossible, but it helps because the model is reading from supplied evidence rather than relying only on patterns learned during training.

Auditability means keeping a record of what the agent did. Which documents did it retrieve? Which tools did it call? What did each tool return? Which step produced the final recommendation? Was a human asked to approve the result?

This matters in fields such as finance, healthcare, law, insurance, and hiring. In those settings, it is not enough for a system to produce an answer. Someone may need to inspect how the answer was reached.

A useful agentic system should leave a trail.

Knowing when to stop

When Agentic Workflows Are Not Worth It#

Agentic workflows are powerful, but they are not always the right choice.

If the task is simple, a direct prompt may be better. Asking for a quick rewrite, a short explanation, a subject line, or a simple brainstorm probably does not need multiple agents, tool calls, and review loops.

Agentic workflows add cost and latency. They also add more places where something can go wrong. A bad search result, faulty tool call, weak summary, or confused coordinator can damage the final output.

The best use cases have several features:

• The task has multiple steps.
• The answer depends on current or private information.
• The work benefits from checking or revision.
• Parts of the task can run in parallel.
• The result is valuable enough to justify extra cost.
• The system can access the required tools and data.
• Mistakes can be detected, reviewed, or escalated.

If most of those conditions are not present, a single well-written prompt may be enough.

Reference

Glossary#

Direct prompting

Asking a model a question and taking whatever it produces in that one attempt.

Agentic workflow

Giving the model a process instead of a single question, so it can plan, use tools, review its own work and try again.

HumanEval

A benchmark of small programming tasks used to test whether a model can write working code.

pass@1

How often a model gets the answer right on its first attempt.

Schema

A fixed format for information, so the model fills in named fields instead of inventing a shape.

API

A defined way for software systems to talk to each other. In a workflow it gives the model a short menu of permitted actions.

Parallelism

Doing several parts of a job at the same time, rather than one after another.

Context window

The amount of text a model can consider at once, including the prompt, the documents, the tool results and its own output.

Coordinator

The agent that decides what work the other agents are given.

Modularity

Building a system from parts that can be replaced independently.

MCP

Model Context Protocol. A standard way for AI systems to connect to tools, files, databases and services.

A2A

Agent2Agent. A protocol for agents built on different platforms to talk to each other.

Grounding

Tying an answer to real sources or tool results rather than to memory alone.

RAG

Retrieval Augmented Generation. Finding the relevant documents first, then asking the model to answer from them.

Hallucination

A confident answer that is not true. It may be fluent and still be unsupported or invented.

Auditability

Keeping a record of what the system retrieved, which tools it called, and which step produced the result.

The takeaway

The Bottom Line#

Agentic workflows make AI more useful because they treat work as a process rather than a single response.

They improve performance by letting models plan, act, review, and revise. They save time by allowing several parts of a task to run at the same time. They are easier to improve because each component, such as the model, search tool, database connection, or review step, can be replaced independently.

The main challenge is no longer only model intelligence. It is whether the surrounding systems are ready. The data has to be accessible. The tools have to be connected. Permissions have to be controlled. The workflow has to be designed so the agent can do useful work and leave behind a record people can inspect.

Used well, agentic workflows move AI from answering questions to completing structured work.

If you want to see these three mechanisms built rather than described, the projects go step by step through the code: Multi-Agent Orchestration builds the coordinator and its subagents, Task Decomposition for AI Agents covers how a large goal gets broken into steps that survive contact with a real model, and Workflow Enforcement and Handoff covers the checkpoints, the prerequisite gates and the escalation path to a human.

Sources

References#

Benchmark scores, forecasts and trial figures below are snapshots verified 2026-07-27 against the cited primary sources. The Gartner predictions in particular are forecasts, not measurements.

Footnotes#

1. Andrew Ng, “Four AI Agent Strategies That Improve GPT-4 and GPT-3.5 Performance,” The Batch, DeepLearning.AI, Mar 20, 2024. deeplearning.ai. ↩

2. Liu, Lin, Hewitt, Paranjape, Bevilacqua, Petroni, Liang, “Lost in the Middle: How Language Models Use Long Contexts,” TACL 12, 2024. arxiv.org. ↩

3. Anthropic, “Introducing the Model Context Protocol,” Nov 25, 2024. anthropic.com. ↩

4. Linux Foundation, “Linux Foundation Launches the Agent2Agent Protocol Project,” Jun 23, 2025. linuxfoundation.org. ↩

5. Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” Jun 25, 2025. gartner.com. ↩ ↩²

6. Boiko, MacKnight, Kline, Gomes, “Autonomous chemical research with large language models,” Nature 624, Dec 20, 2023. nature.com. ↩

7. Insilico Medicine, “Nature Medicine Publication of Phase IIa Results Evaluating Rentosertib,” Jun 3, 2025. insilico.com. ↩

8. Insilico Medicine, “Insilico Initiates Phase III Clinical Trial for Rentosertib.” prnewswire.com. ↩

9. Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” NeurIPS 2020. arxiv.org. ↩