Agentic workflows: a breakthrough in automation?

Business automation in companies is now in a period of transition.
RPA has taken on “doing fixed tasks on someone’s behalf,” and job management systems (hereafter, workflow engines) have taken on “running routine tasks in order.” What has emerged next is an approach that structures and automates “thinking” itself: agentic workflows.

This article goes beyond simply using AI and looks at a shift that leads to redefining the work itself. It explains that shift concretely from the perspectives of work design, PoC adoption, and technical architecture.

The next move: “automating thinking”

In the past, the part of work that was said to “require human hands” was the step of “thinking.”
Ambiguous judgments, forming hypotheses, organizing information: these could not be written down as deterministic rules, and they stayed as tacit knowledge held by specific individuals.

Now, however, the following technical advances are overturning that assumption:

Reasoning ability of large language models (LLMs): summarizing information, forming hypotheses, imitating decision-making
Frameworks such as LangChain and CrewAI: division of roles and collaboration among multiple AI agents
Fast inference with local LLMs: embedding LLMs in business processes with realistic response times becomes possible
Wider adoption of RAG and memory mechanisms: custom LLM behavior grounded in internal company knowledge becomes feasible

Against this backdrop, a new form of work design, in which “AI handles ambiguous work while thinking it through,” is now reaching the front lines of companies.

What does “automating thinking” mean?

“Automating thinking” does not simply mean imitating human output. It means handing the following four elements over to machines as observable and controllable processes:

1. Problem identification (defining the problem space): restructuring “what should be solved”
2. Hypothesis formation (exploring the strategy space): what paths are possible given the constraints and goals
3. Reasoning and selection (reasoning in the semantic space): judging “why this is the best option” based on context and knowledge
4. Execution and feedback (returning results to knowledge): evaluating whether the output is valid and feeding that back into the next attempt

Until now, work has relied on humans holding this process as “tacit knowledge.” Now, LLMs and agents can make these four phases visible, recorded, and reproducible. That is the essence of “automating thinking.”

In this sense, “automating thinking” is not just about making tasks more efficient. It can also be a path for companies to turn their own decision-making structures into “explicit knowledge.”

In other words, as long as judgment that depends on specific individuals remains a “black box,” business processes cannot be automated at scale. Building a process in which AI is made to “think, record its failures, and learn from them” externalizes a company’s intellectual capital and has the potential to lead to large productivity gains and a competitive advantage.

What is an agentic workflow?

Overview and components

Agent: AI that thinks and acts autonomously according to its goal
Tool: a module that connects to the real world, such as external APIs or file operations
Memory: storage that retains past state and conversation history to keep context going
Flow control: controls the flexible flow of a process, including conditional branches, loops, and exception handling

The decisive difference

Layer	Technology	Key characteristics
Operation layer	RPA	Routine GUI operations
Flow layer	Workflow (Airflow, etc.)	Deterministic flow control
Cognitive layer	Agentic workflow	“Automating thinking,” including non-deterministic judgment and trial and error

How will agentic workflows change work?

1. Handing “work design” to AI

Before: humans design the rules
Going forward: AI observes business workflows, forms hypotheses about them, and improves them

Example: an LLM redesigns inquiry-response templates from past history → autonomously improves the knowledge base

2. Making tacit knowledge visible and reusable

Prompts, outputs, the reasoning behind decisions, the tools used, and so on are logged and made reproducible
As a result, “the organization’s thinking process” becomes an asset

3. Achieving “self-improving workflows”

The workflow runs while observing and correcting the LLM’s own behavior
Task accuracy keeps improving without human tuning

Use cases in real work

Department	Example use	Notes
Sales and marketing	Customer research + drafting proposal documents	Web search + PDF summarization + structure suggestions
Customer support	Summarize inquiry → draft a response → improve similar Q&A	Keeps human-in-the-loop
Accounting and legal	Structuring invoices and contracts, plus a review flow	Hybrid of GPT and rule-based processing
SRE	Root cause analysis → generate reproduction steps → propose a recovery	LangChain + tool agent setup
HR	Generate evaluation comments → check context → manager review	Includes checking consistency with evaluation criteria

How should adoption start? From PoC to production

1. Choose suitable work

Non-routine, but with patterns
Humans can evaluate the quality of the output
Logs and prompts can be reused

2. Start small

Insert an LLM into just one step and observe the effect
Make human intervention (review, approval) mandatory at first

3. Make the components visible

Record memory, logs, and prompts
Make explicit which agent did what

4. Evaluate continuously with KPIs

Recall, hallucination rate, review reduction rate, and so on
Human satisfaction and correction rates are also targets for measurement

Requirements from a technical architecture perspective

Why a workflow platform built for agentic workflows is needed

Required capability	Description
Flexible choice of memory	A design that lets you choose a vector DB, graph DB, or key-value DB for each case
Prompt and LLM parameter management	Editing from the UI, version control, parameter settings (temperature, max tokens, etc.)
Multi-agent setup	A structure that coordinates multiple specialized agents (a LangGraph-style representation)
History review and debugging	Essential for non-deterministic execution
Control through RBAC	Resource access control for each agent (external APIs, internal DBs, etc.)

Standardizing LLM integration with MCP, and its potential

For agentic workflows to become deeply embedded in real work, integration between LLMs and external systems (internal DBs, SaaS, tools, and so on) is unavoidable.
A core technology drawing attention here is MCP (Model Context Protocol), proposed by Anthropic.

What is MCP?

MCP (Model Context Protocol) is the technical key that lets agents connect to “the reality inside the company.”

MCP is a standardized communication protocol that lets LLMs access external data sources and capabilities.
Its features include the following:

Defines two-way communication between LLMs and external tools
The LLM requests the information it needs through an MCP client, and the MCP server returns a response using external data and capabilities.
A lightweight specification based on JSON-RPC
It allows general-purpose implementations, and SDKs are available for Python, TypeScript, Java, Go, and other languages.
Conversations can keep state and context
Stateful exchanges are possible, which enables “meaningful conversations” that go beyond simple API calls.

Why does it matter?

With protocols like MCP in place, companies can open their own business systems and APIs to AI agents as things the LLM can “think about.”

Traditional RPA and workflows “repeated what they were told,”
but when AI agents work with external tools through MCP, the following becomes possible:

The LLM consults the customer DB and judges “how this case should be handled”
It examines accounting system history and forms hypotheses for expense classification
It retrieves relevant documents from the internal knowledge base, then summarizes them, compares them, and suggests improvements

Benefits of adopting agentic workflows

Beyond imitating human judgment, it can “externalize the structure of thinking and turn it into an asset”
It can bring reproducibility even to “ambiguous work” that has never been standardized
Starting from collaborative design between humans and AI (human-in-the-loop), advanced co-creative work becomes possible

Risks and limitations in practice

Agentic workflows are a concept with great potential, but they also come with several current constraints and caveats.
For practical adoption, it is important to fully understand the following risks and validate them during the PoC phase.

Hallucinations and factual errors

By their nature, LLMs may output plausible but false information (so-called “hallucinations”).
Especially for “work with a clear correct answer,” such as database contents, contract clauses, or business procedures, the design must assume human review (human-in-the-loop).

Expectation gaps within the organization

When excessive expectations such as “leave it to AI and it will get smarter automatically” spread across the front lines, dissatisfaction with performance and mistakes right after adoption tends to grow.
It is more realistic to see AI agents less as autonomous task executors and more as “interns who are still growing.”

Security and handling of internal data

There will be cases where internal non-public information (documents, DBs, SaaS, and so on) must be given to the LLM as business knowledge or as input for judgment.
In such cases, the following are required:

Running models on-premises or within a VPC
RBAC (permission management) and access logging
Preventing information leaks by validating prompt outputs

By using protocols like MCP to clearly define the “control boundary for what information the LLM is shown,” these risks can be mitigated.

Closing: the next automation starts with “delegating”

From “teaching AI the correct answer for the work” to “letting AI think, then evaluating the result.”
This reversal is the true value of agentic workflows.

Of course, there is no need to replace everything right away. Rather, AI can improve precisely because humans observe and correct it.

Work that depends on specific individuals, knowledge that has never been written down, and decisions made without documentation are common in Japanese companies. These realities are exactly where agentic workflows have the most room to be applied.

Descarty is working to transform the “structure of thinking” in work itself through Dagu, an open-source workflow engine, and Kitewell, business automation software that runs on top of it.

Support for designing and running PoCs with small teams
Building UIs for managing prompts and LLM parameters
Connecting to secure cloud and on-premises infrastructure
Adopting “agentic workflows” that make decisions autonomously

“Externalize thinking.” In the era ahead, that is where work design begins.