AGNTCon + MCPCon Japan brought the agent and MCP community to Tokyo on September 10–11, 2026, and this was an important one for us at NitroStack.
Two members of the NitroStack team were speaking on the first day.
Abhishek Pandit, Founder and CEO of NitroStack, presented Tasks at Scale: How MCP's New Task Lifecycle Powers Long-Running Agentic Systems.
Satyam Soni, together with Rudraksh Karpe, presented Let AGENTS.md Write Itself: Self-Improving Coding Agents, No RL Required!
The two talks were quite different technically, but both came from problems that start showing up when agents move beyond a quick demo and begin doing real work.
Abhishek Pandit on stateless MCP and long-running tasks
Abhishek's session came directly from something we had already dealt with while running agentic systems.
MCP's Tasks API changed.
What had started as an experimental feature evolved into a proper extension. The task lifecycle changed, tasks/list disappeared, and some of the assumptions used to track asynchronous work changed with it.
That matters when an agent is doing something that takes more than a few seconds.
Think about an agent starting a data pipeline, waiting for an approval, processing a large workflow or coordinating several operations across different systems. The HTTP request may be stateless, but the work itself clearly isn't.
That was the core of Abhishek's session.
The newer MCP Tasks model allows a tool call to create a task that can then be tracked through its own lifecycle. Clients can get its current state, update it or cancel it without requiring the MCP transport to maintain a traditional session for the entire operation.
This fits closely with the direction NitroStack has already taken.
We recently wrote about this in more detail in Introducing Stateless MCP to NitroStack.
With the newer stateless MCP architecture, requests no longer need to stay attached to one server session. Each request carries what the server needs to process it, which means another replica behind a load balancer can handle the next request.
For production MCP infrastructure, that's a big improvement.
It makes horizontal scaling cleaner, removes the need for sticky routing in the normal request path, and fits much better with the way modern HTTP infrastructure is already operated.
NitroStack's framework was already heavily request-scoped, so the move toward stateless MCP fits naturally into the architecture we're building.
Long-running work is where stateless MCP gets interesting
Statelessness sounds simple until an agent starts something that needs to continue after the original request is gone.
That's where Tasks become important.
A production agent might kick off:
- a multi-step enterprise workflow
- a long-running research job
- a deployment
- a document-processing pipeline
- an approval process involving a human
- several tool calls running across different systems
The MCP server shouldn't need to keep an HTTP session alive while all of that happens.
The newer task lifecycle gives developers a cleaner way to separate request state from work state.
That distinction was a big part of Abhishek's talk in Tokyo, and it is also becoming increasingly important to how NitroStack approaches production MCP servers.
Satyam Soni on coding agents that improve their own instructions
Satyam's session explored a completely different problem.
Coding agents increasingly depend on files such as AGENTS.md to understand how a repository works: which commands to run, which conventions to follow, how tests should be executed and what mistakes should be avoided.
The problem is that those instructions become stale.
The repository changes, developers discover better ways of doing things, an agent takes the wrong path several times, but the instruction file usually remains untouched.
Satyam and Rudraksh's work explores how that file can improve using the agent's own execution history.
The flow is deliberately controlled.
An agent performs a task. Its trajectory is captured. The run is reviewed for something worth learning. A new instruction can then be proposed, evaluated and promoted only when it passes the required checks.
So instead of blindly allowing an AI model to rewrite its own rules, the system builds an evidence-based feedback loop around them.
That is a much more practical way to think about self-improving coding agents.
A good MCP conversation for NitroStack to be part of
What we liked about AGNTCon + MCPCon Japan was how technical the conversation had become.
The agenda wasn't limited to explaining MCP or showing basic tool calls. People were talking about task orchestration, stateless transports, agent security, evaluation, observability, coding agents and the infrastructure needed to keep these systems running properly.
This is very close to the work happening inside NitroStack.
The NitroStack SDK gives developers a structured framework for MCP applications. NitroStudio gives them a place to build, inspect and test those applications. NitroCloud handles the deployment and production infrastructure around them.
And as MCP itself changes, the framework has to move with it.
Stateless requests, task orchestration, OAuth, observability and production deployment aren't separate conversations for us. They meet inside the same application once somebody tries to ship a serious MCP server.
That's why having both Abhishek and Satyam speaking in Tokyo mattered.
They weren't presenting a generic introduction to MCP. They were talking about specific engineering problems that appear after developers have already started building.
NitroStack in the global MCP ecosystem
AGNTCon + MCPCon Japan also put NitroStack in the room with developers, maintainers and teams working on different parts of the agentic stack.
For a company building infrastructure around Model Context Protocol, that visibility matters.
NitroStack is already being used to build, test and deploy MCP applications, and our focus now is making that development experience hold up as MCP moves deeper into production environments.
Tokyo was another opportunity to show that work publicly.
And having Abhishek Pandit and Satyam Soni both speaking at AGNTCon + MCPCon Japan 2026 made the message fairly straightforward: NitroStack isn't only shipping MCP tooling. The team building it is actively contributing to the technical conversations around where MCP goes next.