{"id":1285,"date":"2026-06-18T19:25:28","date_gmt":"2026-06-18T10:25:28","guid":{"rendered":"https:\/\/www.mindware-jp.com\/en\/?p=1285"},"modified":"2026-06-18T19:25:28","modified_gmt":"2026-06-18T10:25:28","slug":"the-fate-of-ai-agents","status":"publish","type":"post","link":"https:\/\/www.mindware-jp.com\/en\/2026\/06\/18\/the-fate-of-ai-agents\/","title":{"rendered":"The Fate of AI Agents"},"content":{"rendered":"\n

Question for ChatGPT:<\/p>\n\n\n\n

During the development of new media in the 1980s, Videotex was expected to be the leading new medium, but that did not actually come to pass. During the multimedia boom around 1990, interactive media using CD-I and CD-ROM were developed in droves, but they disappeared within a few years. It can be said that this vision was ultimately realized in the form of the World Wide Web (WWW) with the commercialization of the Internet starting in 1994. In the world of technology, it is not uncommon for initial ideas to be realized only two or three generations later. AI itself, having gone through two booms in the 1950s and 1980s, has finally reached a practical level in this third boom. However, not everything we see in the current AI boom will likely endure. I believe agents are at the very core of this, and I would like to sort out what is real and what is an illusion.<\/p>\n\n\n\n

ChatGPT’s response:<\/p>\n\n\n\n

You are exactly right. I think the best way to understand this is through the pattern that \u201cthe idea itself is real, but the first implementation form disappears.\u201d<\/strong><\/p>\n\n\n\n

To state the conclusion first, the core that will remain from the current AI boom is not simply \u201cagents,\u201d but rather an execution intelligence layer<\/strong> that can receive human intent, refer to organization-specific context, use external tools, record its intermediate process, and advance work in a form that humans can verify.<\/p>\n\n\n\n

On the other hand, much of what is now being discussed \u2014 such as \u201cautonomous agents will do all company work on their own\u201d or \u201cif multiple agents gather, advanced intelligence will naturally emerge\u201d \u2014 is largely fantasy.<\/p>\n\n\n\n

1. Historically, what fails first is not the \u201cidea,\u201d but the \u201cform\u201d<\/h2>\n\n\n\n

Videotex had a fairly correct idea: electronic information services for households. However, terminals, communication speed, display capability, content distribution, billing, and standardization were all immature. Britannica describes videotex as an early end-user information system that delivered information to home terminals via telephone lines or cable.<\/p>\n\n\n\n

CD-i and CD-ROM multimedia were similar. The idea of interactive educational, entertainment, and information content was valid. But production costs were high, content could not be easily updated, distribution depended on physical media, and network effects were weak. CD-i was expected in 1991 as an interactive multimedia device for households, but ultimately failed to create a major market.<\/p>\n\n\n\n

Later, with the emergence of the WWW, the same concept of an \u201cinteractive information space\u201d was finally realized through open standards, browsers, URLs, links, search, low-cost publishing, continuous updating, and network effects. Mosaic made it easier for general users to access web pages and image libraries in 1993, and the arrival of Netscape in 1994 popularized the Web.<\/p>\n\n\n\n

In other words, what failed was not \u201cnew media\u201d itself. What failed was a closed, heavy, non-updatable implementation form that could not make users, creators, and distributors move together at the same time<\/strong>.<\/p>\n\n\n\n

2. The same thing will happen with AI agents<\/h2>\n\n\n\n

The current agent boom is probably the same.<\/p>\n\n\n\n

The broad direction \u2014 \u201cAI thinks, investigates, judges, and acts on behalf of humans\u201d \u2014 is real. However, many of the agent products we see today are probably not the final winning form. They are more like the videotex phase or the CD-ROM phase.<\/p>\n\n\n\n

In fact, Gartner stated in 2025 that AI agents and AI-ready data were among the fastest-moving technologies in AI, while also placing them at the \u201cpeak of inflated expectations.\u201d Gartner also predicted that by the end of 2027, more than 40% of agentic AI projects would be canceled due to rising costs, unclear business value, or insufficient risk control.<\/p>\n\n\n\n

McKinsey also notes that although AI adoption is spreading, only a minority of companies have deeply embedded advanced capabilities such as agents into business workflows. In its 2026 AI trust survey, McKinsey identified security and risk as the biggest barriers to scaling agentic AI, with inaccuracy and cybersecurity among the main risks.<\/p>\n\n\n\n

Industrial research also shows that many companies are still at the stage of \u201cAI assistants\u201d or \u201cAI support,\u201d while full-scale multi-agent orchestration remains rare. The central issue is not only capability, but the lack of mechanisms for verifying outputs and safely integrating them into production workflows.<\/p>\n\n\n\n

3. What will remain as \u201creal\u201d<\/h2>\n\n\n\n

I think the following five things will remain.<\/p>\n\n\n\n

1. Natural language interfaces<\/h3>\n\n\n\n

This is unquestionably real.
Humans will no longer directly operate programs, search formulas, or business screens. Instead, they will express intent in natural language. This is a major UI shift after the GUI, the Web, and smartphones.<\/p>\n\n\n\n

However, the chat screen itself may not be the final form. In the future, natural language operation will be embedded into business applications, email, CRM, IDEs, internal wikis, BI, accounting, legal systems, and so on.<\/p>\n\n\n\n

2. Tool-using AI<\/h3>\n\n\n\n

The shift from \u201cAI that answers\u201d to \u201cAI that operates\u201d is real.
Searching, reading files, creating tables, writing code, adjusting schedules, drafting emails, calling APIs \u2014 all of these clearly have practical value.<\/p>\n\n\n\n

However, this will probably settle not as a general-purpose autonomous agent, but as business-specific agents with limited permissions and clearly defined scope<\/strong>.<\/p>\n\n\n\n

3. Context layers and memory layers<\/h3>\n\n\n\n

This is the most important point.
The difference in AI capability will be determined not only by the model itself, but by what context is given to it, what memory is retained, what history is referenced, and what judgment criteria it possesses.<\/p>\n\n\n\n

Current RAG is still mostly search-oriented. In the future, what will matter is not just document search, but a layer that preserves organizational decision history, conceptual structures, policies, constraints, failure cases, and tacit knowledge<\/strong>. This is highly consistent with the direction of ThinkNavi and ConceptMiner.<\/p>\n\n\n\n

4. Human-in-the-loop execution workflows<\/h3>\n\n\n\n

Rather than complete autonomy, what will remain for the time being is a form in which \u201cAI advances the work, and humans approve at key points.\u201d<\/p>\n\n\n\n

Research, summarization, comparison, hypothesis generation, document creation, code generation, checklist creation, task extraction from meeting minutes, and drafting customer response proposals are all strong use cases.<\/p>\n\n\n\n

What matters is not just the AI\u2019s output, but whether the intermediate process, grounds, reference information, reasoning, and change history remain visible<\/strong>.<\/p>\n\n\n\n

5. Evaluation, auditing, and governance<\/h3>\n\n\n\n

If agents execute business work, \u201cverifiability\u201d becomes more important than model performance.<\/p>\n\n\n\n

Research has also pointed out that agents can hallucinate at each stage of task execution, making detection, mitigation, and monitoring necessary.<\/p>\n\n\n\n

In other words, the real market for AI agents is not \u201cclever conversation,\u201d but a business execution infrastructure that can be safely entrusted with work<\/strong>.<\/p>\n\n\n\n

4. What is likely to disappear as fantasy<\/h2>\n\n\n\n

1. The fantasy of the \u201cfully autonomous employee\u201d<\/h3>\n\n\n\n

The idea that AI can behave like a human employee \u2014 understanding vague instructions, grasping context, considering stakeholders, judging risks, handling exceptions, and completing work responsibly \u2014 is still far away.<\/p>\n\n\n\n

AI is extremely strong at partial tasks. But the difficulty of company work lies not only in knowledge, but also in responsibility, priorities, politics, exception handling, tacit agreements, and explanation when things go wrong.<\/p>\n\n\n\n

2. The fantasy that \u201cmulti-agent systems automatically become intelligent\u201d<\/h3>\n\n\n\n

When multiple AIs are given roles and made to hold discussions, the result can look sophisticated.<\/p>\n\n\n\n

However, without shared context, evaluation criteria, role boundaries, stopping conditions, and verification mechanisms, they merely consume a large number of tokens.<\/p>\n\n\n\n

The essence of multi-agent systems is not the number of agents. It is shared memory, role design, task decomposition, evaluation functions, consensus formation, and audit logs<\/strong>.<\/p>\n\n\n\n

3. The fantasy that \u201cprompts alone become products\u201d<\/h3>\n\n\n\n

Prompt collections, wrapper-style chatbots, and general-purpose agent builders may sell in the short term, but they are likely to become commodities in the long term.<\/p>\n\n\n\n

That is because their functions will be absorbed by the model platforms themselves.<\/p>\n\n\n\n

What will remain is not prompts, but business data, proprietary workflows, customer context, evaluation data, execution permissions, and accumulated decision history<\/strong>.<\/p>\n\n\n\n

4. The fantasy of an \u201cagent marketplace\u201d<\/h3>\n\n\n\n

A model that sells general-purpose agents side by side may survive to some extent, but it is unlikely to become the main battlefield.<\/p>\n\n\n\n

That is because value does not arise from the agent alone. It arises when the agent is connected to a specific company, specific business process, and specific data.<\/p>\n\n\n\n

This resembles how many CD-ROM titles were created, but most were eventually absorbed into the Web. Generic agents may also ultimately be absorbed into operating systems, browsers, business SaaS, and LLM platforms.<\/p>\n\n\n\n

5. The real target is not \u201cagents,\u201d but the \u201cagency stack\u201d<\/h2>\n\n\n\n

I see the structure that will remain as follows.<\/p>\n\n\n\n

Layer<\/th>What will remain<\/th>What is likely to disappear<\/th><\/tr><\/thead>
Model<\/td>High-performance LLMs and multimodal models<\/td>Small, undifferentiated proprietary models<\/td><\/tr>
UI<\/td>Natural language operation and AI embedded in business tools<\/td>Mere chat windows<\/td><\/tr>
Memory<\/td>Organization-specific context, decision history, conceptual structures<\/td>Long chat histories left as-is<\/td><\/tr>
Execution<\/td>APIs, tools, and workflow integration<\/td>Demo-style automatic operation<\/td><\/tr>
Management<\/td>Permissions, approvals, logs, audits<\/td>Black-box autonomous execution<\/td><\/tr>
Evaluation<\/td>Testing, evidence checking, reproducibility<\/td>\u201cThe AI said so\u201d<\/td><\/tr>
Business integration<\/td>Embedding into specific business processes<\/td>Standalone general-purpose agents<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

In other words, what is real is not the agent itself, but the infrastructure that makes agency possible<\/strong>.<\/p>\n\n\n\n

6. From the perspective of ThinkNavi and ConceptMiner<\/h2>\n\n\n\n

Seen from this perspective, it would be stronger to position ThinkNavi and ConceptMiner not as \u201cAI agent products\u201d head-on, but as follows:<\/p>\n\n\n\n

ThinkNavi = a workspace where humans and AI think strategically together<\/strong>
ConceptMiner = a conceptual structure, long-term memory, and context-modeling layer that agents can refer to<\/strong><\/p>\n\n\n\n

What is lacking in the world of agents is not simply LLM capability. What is lacking is \u201cstructured context\u201d that answers questions such as:<\/p>\n\n\n\n

    \n
  • What assumptions should decisions be based on?<\/li>\n\n\n\n
  • Which past decisions should be referenced?<\/li>\n\n\n\n
  • Which concepts are close to each other?<\/li>\n\n\n\n
  • Where are the gaps in thinking?<\/li>\n\n\n\n
  • Which hypotheses are weak?<\/li>\n\n\n\n
  • Why did we reach this conclusion?<\/li>\n\n\n\n
  • What should be verified next?<\/li>\n<\/ul>\n\n\n\n

    If ConceptMiner\u2019s GNG+MST, concept maps, structural extraction from chat histories, FactCollector, and LLM Wiki are connected to this layer, it becomes an agent-oriented context infrastructure<\/strong> that is higher-level than ordinary RAG.<\/p>\n\n\n\n

    7. A practical checklist for distinguishing what is real<\/h2>\n\n\n\n

    When evaluating AI-agent-related technologies or businesses, the following questions are useful.<\/p>\n\n\n\n

    Closer to real<\/strong><\/p>\n\n\n\n

      \n
    • Is it deeply connected to the execution of a specific business process?<\/li>\n\n\n\n
    • Does it have proprietary data, proprietary context, and proprietary workflows?<\/li>\n\n\n\n
    • Are human approval points designed into the process?<\/li>\n\n\n\n
    • Are evidence and operation logs preserved?<\/li>\n\n\n\n
    • Can causes be traced when failure occurs?<\/li>\n\n\n\n
    • Does evaluation data accumulate?<\/li>\n\n\n\n
    • Does value remain even if the model changes?<\/li>\n\n\n\n
    • Does the customer feel not that it is \u201cinteresting,\u201d but that time, cost, or risk has been reduced?<\/li>\n<\/ul>\n\n\n\n

      Closer to fantasy<\/strong><\/p>\n\n\n\n

        \n
      • It looks impressive in demos, but there is no explanation of production operation.<\/li>\n\n\n\n
      • It emphasizes \u201cfull autonomy\u201d too much.<\/li>\n\n\n\n
      • It differentiates only through prompts and UI.<\/li>\n\n\n\n
      • Its explanations of auditing, permissions, and security are weak.<\/li>\n\n\n\n
      • It merely lines up many general-purpose agents.<\/li>\n\n\n\n
      • Responsibility in case of failure is unclear.<\/li>\n\n\n\n
      • It seems easily replaceable by a model update.<\/li>\n\n\n\n
      • Customer-specific memory and context do not accumulate.<\/li>\n<\/ul>\n\n\n\n

        8. My conclusion<\/h2>\n\n\n\n

        What will remain from the current AI boom is not simply the AI agent<\/strong>, but the system of context, memory, execution, verification, and auditing that allows work to be entrusted to AI<\/strong>.<\/p>\n\n\n\n

        Just as the essence of 1980s new media was \u201cinteractive access to information from the household,\u201d
        and just as the essence of 1990s multimedia was \u201can interactive knowledge experience integrating text, images, audio, and video,\u201d
        the essence of today\u2019s AI agents is probably this:<\/p>\n\n\n\n

        Transforming human intent into verifiable action in light of organization-specific context.<\/strong><\/p>\n\n\n\n

        In that sense, today\u2019s \u201cagents\u201d are still in the CD-ROM phase.<\/p>\n\n\n\n

        What will truly remain is the equivalent of the WWW: an open and connectable execution infrastructure \u2014 in other words, a common layer through which AI connects to tools, knowledge bases, business systems, and human decision-making processes<\/strong>.<\/p>\n\n\n\n

        <\/p>\n","protected":false},"excerpt":{"rendered":"

        Question for ChatGPT: During the development of new media in the 1980s, Videotex was expected to be the leading new medium, but that did not actually come to pass. During the multimedia boom around 1990, interactive media using CD-I and CD-ROM were developed in droves, but they disappeared within a few years. It can be […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1285","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/posts\/1285","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/comments?post=1285"}],"version-history":[{"count":1,"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/posts\/1285\/revisions"}],"predecessor-version":[{"id":1286,"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/posts\/1285\/revisions\/1286"}],"wp:attachment":[{"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/media?parent=1285"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/categories?post=1285"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mindware-jp.com\/en\/wp-json\/wp\/v2\/tags?post=1285"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}