自作言語で AI エージェントをオーケストレーションしよう！Katari, a custom language (v0.1.x) designed for AI agent orchestration, is…

A developer has begun building Katari, a custom programming language aimed specifically at writing AI agents and orchestration logic, and has published an early look at the project alongside a working Discord bot that demonstrates what the language can do today. The project is currently at version 0.1.x, an early stage that signals the language is still taking shape rather than being production-ready. For teams experimenting with autonomous agents, a language designed from the ground up for orchestration is a notable direction, because much of today's agent tooling is layered on top of general-purpose languages rather than built around the problem.

The clearest way to understand Katari's intent is through the demonstration project. The developer built a Discord bot in the language that performs several ongoing tasks: periodic email monitoring with reflection into a calendar, recurring scheduled work, and web monitoring. Notably, the bot can start and stop those monitoring routines on its own initiative, rather than requiring a human to trigger each run. It also handles task management, maintains memory, and includes a concept the author refers to as "Soul," along with the use of external tools such as image editing utilities. Taken together, this describes an agent that can watch external signals, act on schedules, manage its own state, and coordinate multiple tools within a single runtime.

The project is documented at katari-lang.dev, with source code hosted at github.com/katari-lang/katari. The accompanying article is described as covering the language specification and future plans, so readers looking for concrete syntax and design decisions will find the primary material there. Because the language sits at v0.1.x, its feature set, interfaces, and semantics are likely to change, and anyone evaluating it should treat the current release as a preview intended to communicate direction and gather feedback rather than a stable foundation for long-lived systems.

Katari is filed under the Model Context Protocol category, and its tags include mcp and mcp-server">mcp-server, which places it within a fast-moving area of the agent ecosystem. MCP is an open standard, introduced by Anthropic, that defines how AI applications connect to external data sources and tools through a consistent interface. In practice, an MCP server exposes capabilities—files, APIs, calendars, and similar resources—that a model or agent can call in a standardized way. A language positioned around orchestration and MCP appears designed to make it easier to wire these tool connections into agent behavior, though the exact depth of Katari's MCP integration is best confirmed from the repository and the language's own documentation.

For context, orchestration has become one of the central challenges in building useful agents. It refers to coordinating multiple steps, tools, and decisions—deciding what to call, when to call it, how to persist memory across turns, and how to recover from failure. Several established frameworks tackle this in existing languages, including LangChain and LangGraph, Microsoft's AutoGen, and CrewAI, most of which are Python or JavaScript libraries. Katari's approach differs by proposing a dedicated language rather than a library, which could allow orchestration concepts—scheduling, monitoring, memory, and tool use—to be expressed as first-class constructs. Whether a purpose-built language proves more ergonomic than framework-based approaches is an open question that will depend on its design and adoption.

The features highlighted in the demonstration map closely to common agent needs. Persistent memory addresses the fact that language models are stateless between calls, so agents require an external store to remember prior context. Scheduled and event-driven monitoring reflects the shift from chat-style, request-response agents toward long-running services that observe the world and act proactively. The ability to autonomously begin and end monitoring tasks hints at a control model where the agent manages its own workload, which is powerful but also raises the usual considerations around predictability, cost, and safeguards.

As an early-stage, open-source effort from an individual developer, Katari is worth watching more as an indication of where agent tooling may head than as a finished platform. Readers interested in the specification, the Discord bot implementation, or contributing feedback can follow the project through its website and GitHub repository, where future releases will show whether the language matures beyond its current 0.1.x state.