Three modes
FIM Agent operates in three modes, determined by how the agent is deployed and used:| Mode | What it is | Delivery | Example |
|---|---|---|---|
| Standalone | General-purpose AI assistant | Portal | Chat, search, code execution, knowledge base Q&A |
| Copilot | AI embedded in a host system | iframe / widget / embed | ”Finance Copilot” embedded in ERP web UI |
| Hub | Central cross-system orchestration | Portal / API | Agent queries ERP, checks OA approvals, notifies via Lark |
Mode details
Standalone (0 connectors)
The default mode. FIM Agent works as a full-featured AI assistant:- Built-in tools: web search, Python execution, calculator, file operations, shell commands
- Knowledge base with RAG (PDF, DOCX, Markdown, HTML, CSV)
- Dynamic DAG planning for complex multi-step tasks
- Real-time streaming with DAG visualization
Copilot (embedded)
Embed FIM Agent into a host system’s web UI. The agent works alongside users in their familiar interface — no context switching required. Copilot mode can use multiple connectors (e.g., the host system’s DB + a notification service). Examples:- Finance Copilot: Connected to Kingdee (金蝶) via DB connector → query financial statements, generate analysis reports
- Contract Copilot: Connected to contract management system via API connector → search contracts, extract clauses, assess risk
- HR Copilot: Connected to HR system via API connector → query employee info, generate statistics
Hub (central orchestration)
The Hub is a standalone portal (or API) that serves as the central intelligence layer. It’s not embedded in any single system — instead, it connects to all of them. Users access it through the Portal UI or API. Examples:- “Check overdue contracts in CRM, cross-reference with ERP payments, notify finance team on Lark”
- “When OA approval completes, update contract status in CRM and log to audit database”
- “Query sales data from Salesforce, generate forecast using business DB, email summary to management”
Delivery methods
| Delivery | Description | Typical mode |
|---|---|---|
| Portal (Web UI) | Built-in Next.js interface | Standalone, Hub |
| API (headless) | HTTP/SSE endpoints (/api/execute, /api/stream) | Hub (programmatic access) |
| iframe / Embed | Injected into host system pages | Copilot |
Execution engines (internal implementation)
Under the hood, FIM Agent provides two execution engines:| Engine | Best for | How it works |
|---|---|---|
| ReAct | Single complex queries | Reason → Act → Observe loop with tools |
| DAG Planning | Multi-step parallel tasks | LLM generates dependency graph, independent steps run concurrently |
Why no traditional workflow engine
FIM Agent deliberately does not build a drag-and-drop workflow editor. This is a strategic choice:- Workflows already exist elsewhere. Enterprise clients’ fixed processes (approval chains, audit flows) live in their OA, ERP, and legacy systems. They need AI that connects to those systems, not another workflow editor.
- Dynamic DAG covers the flexible case. For tasks not pre-defined, LLM-generated DAGs adapt at runtime — no human pre-design required.
- Existing capabilities compose into fixed pipelines. Scheduled Jobs (planned) trigger a DAG agent with a fixed prompt; the DAG plans the steps; Connectors bridge to target systems. The combination equals a static pipeline — but more flexible, because the LLM adjusts its plan based on data it encounters.
- Connector = API call. Complex workflow operations (transfer, reject, escalate) are the target system’s responsibility. From the connector’s perspective, each operation is just an HTTP request with parameters. FIM Agent calls the API; the target system manages the state machine.