![AI Agents Pioneering the Future | A New Horizon of Autonomous AI x Human Collaboration Drawn by 2026's Frontline Players [SusHi Tech Tokyo 2026]](/images/columns/ai-agents-pioneered-future-2026/cover.png)
Hello, this is Hamamoto from TIMEWELL.
"By 2028, there will be 1.3 billion AI agents in the world." With that single line from Microsoft Japan President Miki Tsusaka[^2], the timeframe of the entire session was set. By my reckoning, 2028 is essentially tomorrow. So what is actually happening on the ground in Japan as we head into that tomorrow?
This panel, held at SusHi Tech Tokyo 2026[^1], featured Microsoft Japan President Miki Tsusaka, Telexistence CEO Jin Tomioka[^3], Kakehashi CEO Takafumi Nakagawa[^4], and — bringing the investor's lens — Cathy Matsui, GP at MPower Partners. With physical AI, in-house enterprise AI agents, healthcare AI, and an investor view all on stage, the panel covered the implementation frontier of the AI agent era from multiple angles.
Summary: Three Takeaways from the Session
- Within one month of the Tokyo Metropolitan Government Vice Governor introducing Copilot, 1,000 AI agents were generated inside the metropolitan government. 94% of Nikkei 225 companies are using the Copilot suite. The numbers themselves confirm that the AI agent era has arrived.
- After eight years of trial and error, Telexistence's Tomioka has landed in retail and logistics. Robots are now operating in three of Japan's major convenience-store chains, can be set up in five hours, and run on a hybrid model of 99% computer-vision automation plus 1% remote operator support.
- The success drivers for AI agent deployment break down to "10% algorithm, 20% IT infrastructure, 70% human factors." The depth of an organization's understanding of its people — not its technology — decides who wins.
SusHi Tech as the Convergence Point for AI Agent Implementation
SusHi Tech Tokyo 2026, held at Tokyo Big Sight from April 27 to 29, 2026, is one of Asia's largest innovation conferences. Within this year's "AI and Robotics" theme, AI agents were one of the hottest topics. This session became a venue where unfiltered, on-the-ground reports from real implementations were exchanged.
For TIMEWELL, AI agents are a core component of what we call "infrastructure for challenge." This session was a valuable chance to validate our direction. The room held about 1,000 seats and reached capacity early — a tangible sign of how high the interest in AI agents has become.
For the conference-level highlights, see SusHi Tech Tokyo 2026 Keynote Report.
The Tokyo Government Shock — 1,000 Agents in One Month
The number Tsusaka revealed was striking. Within one month of the Tokyo Vice Governor introducing Microsoft Copilot, 1,000 AI agents were generated inside the metropolitan government. 1,000 in a month — that is evidence of AI agents springing up simultaneously from the edges of the organization.
| Metric | Value | Comment |
|---|---|---|
| AI agents generated inside the Tokyo government | 1,000 in one month | After Vice Governor's Copilot rollout |
| Copilot adoption among Nikkei 225 companies | 94% | Standard equipment for large enterprises |
| GitHub Copilot developer productivity gain | 50% faster coding | Repetitive work automated |
| GitHub Copilot code output | 3x increase | Job satisfaction also rose |
Asked about the concern that "code quality will drop," Tsusaka was unambiguous. GitHub Copilot handles repetitive work, while developers shift toward more creative and strategic tasks. This is not just automation — it is a qualitative change in the nature of the work itself.
Tsusaka's Preview — Agent 365
Tsusaka also previewed the "Agent 365" suite that Microsoft was set to announce the following month. It is a new product for managing in-house agents securely, designed as the foundation for an era in which a single enterprise runs hundreds or thousands of agents.
"AI agents are the new employees of the enterprise," Tsusaka said. Just as new hires need onboarding, permission management, and performance review, AI agents need the same governance. Agent 365 is positioned to be exactly that infrastructure.
Interested in leveraging AI?
Download our service materials. Feel free to reach out for a consultation.
Telexistence's Tomioka — Eight Years of Struggle in Physical AI
Tomioka's presentation distilled eight years of trial and error in deploying physical AI. Since founding the company in 2017, he spent eight years searching for "the right application." The answer he eventually arrived at was the retail and logistics sector.
The current results are concrete. Robots are already running in three of Japan's major convenience-store chains, with a setup time of just five hours. Computer vision automates 99% of operations, while the remaining 1% of complex cases is handled by remote operators in the Philippines — a hybrid model.
The fundamental challenge of physical AI, in Tomioka's framing, is "lack of training data." Language models have an internet's worth of text. But there is no public corpus describing how robots should move in the physical world — you have to collect that data yourself. That is why Telexistence runs a remote-operations center in the Philippines and is accumulating human intervention data through it.
"Secure gross margins of 35% or higher." Tomioka returned to this number repeatedly. A physical AI business does not work economically without a certain scale. He estimates Japan's retail and logistics sectors offer roughly 150,000 person-equivalents of latent labor demand, and Telexistence's strategy is built on closing that enormous labor gap.
Kakehashi's Nakagawa — A Platform Covering 25% of Japan's Pharmacies
The scale Nakagawa described for Kakehashi made an impression. Kakehashi covers 17,000 pharmacies nationwide — over 25% of all pharmacies in Japan — and holds patient interaction data on more than 30 million people. The company is already a piece of Japan's healthcare infrastructure.
Their new product, "AI Musubi Record," places microphones at the pharmacy counter to automatically record and analyze conversations between patients and pharmacists. By starting from cutting record-creation time and simplifying the daily work of pharmacy staff, Kakehashi takes a bottom-up approach that sidesteps psychological resistance on the front line.
This is an extremely smart strategy. If AI is presented as "something that takes your job away," resistance from the field becomes intense. But if it is introduced as "something that helps you with the tedious documentation," acceptance is completely different. Nakagawa has demonstrated, through real operations, that the social deployment of AI agents is decided less by technology than by the quality of the introduction design.
Cathy Matsui — The "10x Agent" Question Investors Are Asking
Matsui's perspective had the sharpness only an investor brings. "Every startup now faces the question: when a 10x more powerful agent shows up, can your business survive?"
This is the question every SaaS startup needs to engage with. In the AI agent era, the following business models will all change at the foundation.
| Traditional model | Change in the agent era |
|---|---|
| SaaS flat-rate subscriptions | Shift toward billing on agent processing volume |
| Customer success | Domains automated by agents will expand |
| Sales cycles | A new form of "agent-to-agent" negotiation emerges |
| Single-vendor lock-in | Multi-agent compositions become the norm |
Matsui also stressed a positive angle for Japan. "Japan's population decline and aging act as a 'forcing function,' raising the urgency of AI adoption." The fact that this headwind can be turned into a tailwind is, I believe, an advantage Japan has that other countries do not.
My Take — The "10-20-70 Rule" of AI Agents
The "10-20-70 rule" that surfaced midway through the session hit me especially hard. The success drivers of AI agent deployment break down to: 10% algorithm, 20% IT infrastructure, and 70% human factors (organization, culture, process).
This matches almost exactly what we see at TIMEWELL day-to-day. The reason most enterprises fail at AI deployment is not that the technology is bad. It is underinvestment in the "human factors" — organizational change, redesign of workflows, and the psychological adaptation of employees.
Nakagawa's bottom-up approach, Tomioka's eight years of iteration, Tsusaka's Agent 365 governance layer — every one of these is an attempt to crack that 70%. Whether AI agents win or lose comes down to the depth of understanding of organizations and people, not the technology.
The Human Role Becomes "Judgment, EQ, Relationship Building"
Another important point Matsui emphasized: in an era where AI handles task automation, what remains for humans is judgment, emotional intelligence (EQ), and the ability to build relationships.
This is an important message for educational institutions. Talent development going forward needs to shift across three areas.
- Judgment: decision-making in complex and ambiguous conditions
- EQ: the ability to read and empathize with the emotions of others
- Relationship building: creating value through trust, cooperation, and negotiation
Both university education and corporate training need to respond to this shift. The transition is from "knowing the right answer" to "creating the right answer." In my work as a specially appointed associate professor at Shinshu University, I sense that this transition has quietly already begun.
The Need for Unemployment Response and Skill Shifting
Another important theme Matsui raised, in her capacity as an investor, was the need to run national-level unemployment response and skill-shifting programs in parallel.
When AI automates many routine tasks, the people who currently do those tasks need infrastructure that lets them move smoothly into the next role. This is not a problem the market or individual companies can solve on their own — it is a national agenda. I believe Japan's unemployment insurance system and vocational training system urgently need to be redesigned for the AI era.
The Telexistence Model Suggests a "Japan x Global Talent" Future
The Telexistence model — robots operating in Japan, humans remote-piloting from the Philippines — is a new employment model that "covers Japan's labor shortage with global talent."
This is a different shape of labor globalization from immigration policy. People do not need to physically relocate to Japan; they can support Japanese on-site operations remotely. The structure of "filling labor demand born in Japan with talent from anywhere in the world" should become an important component for sustaining Japan's competitiveness in the long run.
AI Agent Adoption Models in Financial Institutions
After Tsusaka's session, an interesting number came up from another speaker. Within Japan's major banks, Microsoft Copilot deployments are expected to expand more than threefold from 2025 into 2026. Adoption is particularly strong in legal and compliance functions, with one bank reportedly achieving a 60% reduction in contract-review time.
The financial industry is heavily regulated and is often viewed as slow to adopt AI. But in reality, its mix of high-volume repetitive work and large quantities of standardized documents is a strong fit for AI agents. Using AI while staying compliant — that balance, designed well, is exactly where Japan's financial institutions can build competitive advantage.
AI Agents Trigger a Fundamental Rethink of Organizational Design
Once AI agents are deployed at scale, organizational design itself changes. The traditional "department head — section manager — staff" pyramid moves closer to a flatter structure of "human supervisor — fleet of AI agents."
A single human supervises 10 or 20 AI agents, and those agents execute the operational work. In that structure, the number of mid-level managers declines, while the scope of authority for an individual expands. Flatter hierarchy, faster decision-making, more direct accountability for results — all of these advance simultaneously.
At TIMEWELL, we are getting ahead of this shift and developing talent-development programs for working alongside AI agents. Leadership in the AI era shifts from "managing subordinates" to "designing agents."
A Quick Map of the 2026 Agent Stack
Because the panel referenced several frameworks without deeply contrasting them, I want to add a practitioner's quick map of the agent stack as it actually looks in production today.
Anthropic's Claude Agent SDK has gained traction inside enterprises that prioritize integration safety, tool-use reliability, and longer effective context windows for multi-step reasoning. Its strength is the combination of strong instruction-following and a relatively conservative posture around side effects.
LangChain's LangGraph has emerged as the orchestration layer of choice when teams need explicit control flow — branching, retries, parallel fan-out — across multiple model calls and tool invocations. For complex workflows that must be auditable, LangGraph's graph abstraction maps cleanly onto how operations teams already think about processes.
Microsoft AutoGen and CrewAI both occupy the "multi-agent collaboration" space, with AutoGen leaning toward research-grade flexibility and CrewAI leaning toward role-based templates that ship quickly. Mastra, while newer, is gaining attention in TypeScript-heavy stacks where teams want a tight integration with their existing Node.js infrastructure.
The pattern I keep seeing in real deployments is not "pick one of these and standardize." It is "pick the orchestration layer first, then mix the model providers underneath as the cost-quality frontier moves." This is exactly the architecture that Cathy Matsui's "10x agent" question pushes you toward — keep the orchestration stable, swap the underlying intelligence aggressively.
The Hidden Cost Curve: Long-Running Agents and Token Economics
One of the under-discussed risks in the agent era is cost. A traditional API call charges you for one round trip. An agent that thinks, calls tools, reflects on the result, and iterates can easily multiply that cost by 10x or 50x depending on the design.
In TIMEWELL's own deployments, we have learned to design with three guardrails. First, an explicit budget per task — both in tokens and in wall-clock time — that the agent cannot exceed without escalating to a human. Second, aggressive use of prompt caching for stable system prompts, which materially reduces marginal cost as usage scales. Third, a clear separation between "high-quality model for reasoning" and "cheaper model for execution," chained together by the orchestration layer.
Without these guardrails, agent pilots routinely deliver impressive demos and then quietly fail the unit-economics review. With them, the same workloads can run in production at margins that actually make sense.
Trust, Auditability, and the Boundary of Authority
A theme that ran underneath every speaker's remarks, even when not stated explicitly, was trust. An agent that can act in the world is an agent that can also act incorrectly in the world. Designing the boundary of its authority is the deciding factor in whether enterprises adopt it broadly.
The pattern that is converging across mature deployments has three layers. The first is least-privilege scoping: an agent only gets access to the data and actions that its specific job requires, with role-based access control enforced at the tool level rather than relying on prompt-level instructions. The second is a structured audit log that records every tool call, every decision branch, and the reasoning trace that led to it, in a format that compliance teams can actually review. The third is a clear escalation path — humans-in-the-loop for high-stakes decisions, with clear criteria for what triggers escalation.
Kakehashi's bottom-up design implicitly embodies this. The agent records and summarizes; the pharmacist verifies and acts. Telexistence's hybrid model is the same idea expressed in physical AI. The 99% the robot handles is by design within its scoped authority. The 1% it cannot handle is escalated to a human operator who has both the context and the authority to resolve it.
This pattern — scoped authority plus auditable trace plus structured escalation — is, I think, the actual blueprint for enterprise-grade AI agents. Without it, agents stay in the demo phase. With it, they move into operations.
Closing Thought — What I Want TIMEWELL to Be in This Era
Closing on a personal note: I left this session more convinced than before that TIMEWELL's role is to walk alongside individuals and organizations as they redesign their work with AI as a partner.
The 1.3 billion agents Tsusaka projected for 2028 will not arrive evenly. They will concentrate in the organizations that take the human side of the work seriously — that invest in the 70% before they invest in the 10%. Our job is to make sure that the people we work with are in that group, not the group that gets left behind because they bet on the technology and forgot the people.
If that resonates with how you are thinking about the next two years, I would be glad to talk.
Conclusion — Five Things to Prepare for the AI Agent Era
Distilled from this session, here are five things enterprises should prepare for the AI agent era.
| # | Preparation item | Owner |
|---|---|---|
| 1 | Agent management infrastructure (an integrated platform like Agent 365) | IT / CIO |
| 2 | Bottom-up rollout strategy (minimize psychological resistance on the front line) | Business unit head |
| 3 | Strengthening human judgment and EQ (focus on the work that remains for humans) | HR / L&D |
| 4 | Investing 70% in organizational design (prioritize human factors over technology) | Executives |
| 5 | A collaboration model with global talent (a Telexistence-style hybrid design) | Executives / HR |
For TIMEWELL itself, this is the moment to redesign our "infrastructure for challenge" with AI agents as a base assumption. I want to bring the words of these four leaders — Tsusaka, Tomioka, Nakagawa, and Matsui — back into our own decision-making.
The era of 1.3 billion agents in 2028 is just over the horizon. How Japanese companies and individuals prepare for that era is the question. After hearing this session, I am again convinced that TIMEWELL's role is to walk alongside individuals on the journey of redefining their own work with AI as a partner.
A Note from TIMEWELL
If you are working through AI agent introduction design, organizational design, or building a talent-development program, our AI consulting service WARP offers individual advisory engagements. You can start with a 30-minute online consultation.
Related Articles
- SusHi Tech Tokyo 2026 Keynote Report
- Green Intelligence and the Circular Future [SusHi Tech Tokyo 2026]
- Implementing AI Agents with the Claude Agent SDK
References
[^1]: SusHi Tech Tokyo 2026 official site. https://sushitech-startup.metro.tokyo.lg.jp/ [^2]: Microsoft Japan. https://www.microsoft.com/ja-jp/ [^3]: Telexistence official site. https://tx-inc.com/ [^4]: Kakehashi official site. https://www.kakehashi.life/