EnterpriseAgentFramework, published as ReachAI, is a Java platform for adding agents to business systems that already exist — office automation, human resources, procurement, customer relationship management, ticketing, and contracts. An assistant is embedded in the existing pages, so staff query data, fill forms, and complete tasks in natural language without switching applications. Requests run under the signed-in user's identity and permissions, page navigation and forms run through explicitly registered page actions, and data access still goes through the original backend, which keeps final authorisation.
The adoption path has six stages: connect the system to the runtime layer; scan pages, routes, components, and interfaces into a capability map; analyse selected pages read-only and propose at most three modification candidates, each with confirmed facts, technical inferences, open questions, code references, and acceptance criteria; hand the work to an AI coding tool as a standardised task; publish capabilities, workflows, and agents; and verify against real business results.
Business logic is registered as capabilities. A JDK 8 compatible SDK and Spring Boot 2 starter register a project, its instances, and annotated methods carrying a name, parameters, domain, side-effect level, and required roles; the starter also scans Spring MVC endpoints and reports heartbeats and signatures. The platform stores capability snapshots, computes field-level differences, and lets a developer apply or ignore each change into a capability catalogue. Systems that cannot be modified are inventoried by scanning instead.
At run time a supervisor built on AgentScope handles intent understanding, planning, and workflow selection, then a runtime executes a version-pinned graph specification. Available tools are filtered by capability allowlist, tool access control list, risk level, and page intent, and side-effect levels decide what may proceed: reads run automatically, page actions need an explicit page intent, writes need one user confirmation, and irreversible operations are refused by default. Planning, tool calls, guard decisions, timings, and errors are written to replayable traces.
The platform runs as five services, including a model gateway for chat, embedding, and rerank models and a knowledge service covering knowledge bases, files, and retrieval. Capabilities and published workflows can be exposed as Model Context Protocol tools, and an agent-to-agent gateway makes registered tools and remote agents reachable from other systems under the same governance rules. The stack is Java 17, Spring Boot 3.4.5, Spring AI, AgentScope 2.0.0, LangGraph4j, a Vue 3 front end, MySQL 8, Redis 7, Milvus 2.4, and optional Docling. The project places itself next to Dify, which centres on creating AI applications, and OpenClaw, which drives personal cross-application agents.
Features
- Embedded assistant: a chat entry point and page bridge run inside the existing pages
- Explicit page actions: filters, navigation, reads, and form operations are declared actions, not inferred from screenshots
- Capability SDK: annotated Java methods and Spring MVC endpoints register as governed capabilities with roles and side-effect levels
- System scanning: pages, routes, components, and interfaces are inventoried into a capability map
- AI coding handoff: tasks are packaged with repository context, constraints, and acceptance conditions for external coding tools
- Workflow execution: published workflows run a version-pinned graph specification, selectable by the agent as tools
- Governance policy: allowlists, tool access control lists, risk levels, and confirmation rules gate every call
- Tracing: decisions, tool calls, timings, and errors are recorded for replay
- Open protocols: capabilities and workflows publish as MCP tools, and remote agents connect through the gateway
