Agentican brings config-first multi-agent framework to Java

The author, a former Red Hatter, open-sourced the Agentican Framework after spending a year building what they describe as an "AI workforce platform" — a system where agents act as colleagues to delegate complex, multi-step tasks to, rather than tools to prompt. Their central frustration with existing Java agent frameworks like LangChain4j and Embabel was their code-first design: agents are annotated classes or interfaces, workflows are built programmatically, and everything is coupled to the application that runs it. This means every service that needs a Market Analyst defines its own, and every workflow that involves one becomes a new class hierarchy, leading to duplication as the system scales.

Agentican's answer is to decouple agent definitions from applications entirely.

Agentican's answer is to decouple agent definitions from applications entirely. Agents, skills, and plans live in repositories — in-memory or persistent — and are accessed by the framework at runtime rather than compiled into the application. Tools are managed externally via MCP and platforms like Composio. The framework exposes extension points for custom code steps, tools, agent loops, and event listeners, but treats orchestration as its own responsibility. At the simplest level, a few lines of code invoking `AgenticanRuntime.builder()` with an Anthropic API key is enough for the built-in Planner to read a task description, generate the necessary agents and skills, and execute a plan. In production, agents and plans are defined in `application.yaml` — including LLM config, agent roles, and model assignments like `claude-sonnet-4-5` — and injected into services via `@AgenticanPlan`, reducing orchestration to near-codeless embedding.

Key facts

1. 01The Agentican Framework was open-sourced by a former Red Hatter as a multi-agent framework for Java.
2. 02Agentican stores agents, skills, and plans as repository artifacts independent of any application, enabling reuse across services.
3. 03Existing Java frameworks like LangChain4j and Embabel are code-first: agents are annotated classes and workflows are built programmatically.
4. 04The built-in Planner can read a plain-text task description and autonomously create agents, choose tools, build a plan, and execute it.
5. 05In production, agents and plans are defined in YAML config and injected via `@AgenticanPlan`, requiring minimal application code.
6. 06The framework integrates with MCP and platforms like Composio for external tool access.