What's New in AI Fusion 3

AI Fusion 3 is a re-architecture of the platform. The infrastructure, the agent model, and the development experience in Studio have all changed. Observation keeps its concepts but becomes agent-aware.

This article summarizes the changes and what they mean if you are moving from AI Fusion 1.x or 2.x. The documentation of earlier versions remains available in the legacy section.

At a Glance

Area	AI Fusion 1.x / 2.x	AI Fusion 3
Agent definition	Broadway agentic flows built from actors such as `Chat`, `ChatPromptBuilder` and `LLMAgent`; orchestrator, planner and refiner flows.	Declarative agents: an `agent.yaml` file plus folders for skills, sub-agents, references, evals and observation, inside a Data Product (`<LU>/AI/Agents/<agent>/`). The runtime runs the reasoning loop.
Runtime entry point	The `chat` web service and the `Chat` actor.	The streaming `agent` web service (Server-Sent Events), the `aifusion.Agent` flow and the `AgentInvoke` actor.
Knowledge and procedures	Vector-DB pipelines for KB, procedures and plan examples.	Markdown reference packs under `References/`, searched by a built-in reference-search sub-agent; the `Procedures` pack guides plan-then-execute mode. No vector database is required.
Data access	Custom SQL-generation and describe-domain flows.	The built-in `query` tool over the entity micro-database, guided by schema skills that can embed the live LU schema.
Development environment	Broadway editor and the Chat Playground web app.	The AI Fusion - Studio extension: AI Explorer, form-based Agent Editor, Preview chat with Trace, Eval Editor and AI-assisted agent building.
Evaluation	The web Evaluation workspace, app-scoped suites in MTables, `aifusion.evaluatorPipeline`.	Per-agent suites in `Evals/*.csv`, authored and run in Studio, executed by the Pipeline extension. New performance dimensions (tokens and response time).
Observation	Global signals and widgets.	Core signals and widgets plus per-agent `signals.yaml` and `widgets.yaml`; agent selector on the dashboard; feedback filters; conversation navigation.
Packaging	One `aifusion` extension including demos and vector pipelines.	Core `aifusion`, `aifusion-studio`, separate LLM connector extensions, optional `mcp` and `aifusion-demo`.

New Capabilities

Declarative agents — define role, rules and capabilities in YAML; no flow orchestration to maintain. See Agents Overview.
Sub-agents and delegation — split work across specialist agents, use them as helpers (task) or hand them the whole turn (delegate), and link agents from other Data Products. See Sub-Agents and Delegation.
Skills — on-demand Markdown instructions and schema descriptions, loaded only when needed to keep prompts small. See Skills.
Planning and reference search — plan-then-execute mode guided by written procedures, with optional user approval, and cited answers from reference documents. See References and Planning.
Hooks — Broadway flows that run before the user message, after the answer, around tool calls and on errors. Their inputs define the agent's inputs. See Hooks and Agent Inputs.
Response guidelines and constraints — house style per turn, and non-negotiable rules that survive delegation.
Interactive turns — the agent can ask the user structured questions (ask_user), show a live to-do list, and render charts, tables and links (widgets).
Context management — automatic context compaction for long conversations, prompt caching, configurable reasoning (thinking) levels from off to max.
MCP tools — attach any MCP server, including Fabric Data Products exposed over MCP, as agent tools.
AI-assisted development — Studio AI can build an agent from a Data Product schema, generate test cases, and diagnose a bad answer from its trace.

Removed or Replaced

The Chat actor, the chat web service and the actors ChatPromptBuilder, ChatSessionParams, SetChatSessionParam, WriteToOutputStream, PrependParamsToStream and SaveState were removed (deprecated in 3.0.0, deleted in 3.0.4). Use the agent web service or the AgentInvoke actor instead.
The vector ingestion pipelines, the chunker and the domain demos are no longer part of the core extension. The banking demo ships in the aifusion-demo extension.
.agent files and earlier agent keys (input_params, content_flows, hooks.fabric_get, nested tools: {built_in, custom}, search_reference.dir) are obsolete. Agent inputs are now inferred from the hook flows.
The web Evaluation workspace, aifusion.evaluatorPipeline and MTable-based suites are replaced by per-agent eval suites in Studio. AI-assisted "Copilot" and "Autopilot" test generation are replaced by the Studio AI test generator.
In the insights (assurance) database, ACTVSESSION.app_id was renamed to agent_name and customer_id was replaced by the INSTANCES table. The aifusion deploy migrates existing tables.

Moving from AI Fusion 2.x

AI Fusion 3 is not a drop-in upgrade for 2.x implementations. Agents built as Broadway agentic flows must be re-implemented as declarative agents. A typical migration path:

Install the AI Fusion 3 extensions — see Installation and Setup.
For each 2.x application, create an agent in the Data Product that holds the entity data (AI Explorer → New Agent…).
Move the system prompt and business rules into system_prompt, response_prompt and constraints.
Replace the customer-profile / data-loading logic with a step_start hook flow that loads the entity and declares its inputs (for example customerId).
Replace SQL-generation flows with the built-in query tool and schema skills; keep business operations as flow tools.
Convert KB and procedure documents to Markdown reference packs (References/KB, References/Procedures).
Recreate evaluation suites by saving conversations from the Studio Preview as test cases, or generate them with Studio AI.
Move custom Observation tags into the agent's Observation/signals.yaml.

The Agent Builder in Studio AI can scaffold most of steps 2–5 from the Data Product schema.

Next article: Architecture