AI Fusion Studio is the development environment for AI Fusion agents inside K2view Web Studio. Because an AI Fusion 3 agent is a set of files inside a Data Product (LU) — an agent.yaml, skills, reference documents, sub-agents, evaluation suites and observation configuration — you build it the same way you build the rest of your Fabric project: as files in your project tree. AI Fusion Studio adds the tooling that makes those files easy to create, edit, test and evaluate without leaving Studio.
Note: The examples and screenshots in this section use the banking demo (
customer_bank›banking). See About the Examples.
The tooling is delivered by the AI Fusion - Studio extension. It provides:
agent.yaml and sub-agent files. See Agent Editor.The extension is published on K2exchange as AI Fusion - Studio (version 1.0.2 at the time of writing) and requires Fabric 8.5.2 or later. Install it from the Extensions view of Web Studio, as described in K2exchange. The extension contributes no settings of its own — it talks to the Fabric server that hosts your Studio session.
Installing the extension also adds two Studio AI skills to your project, under .agents/skills/: agent-builder and researcher. They power the AI-assisted features described in AI-Assisted Agent Development. Commit them to Git together with the rest of the extension's files.
For the full list of AI Fusion prerequisites (the runtime extension, the LLM interface and the insights database), see Installation and Setup.
AI Explorer and the editors work on a fixed folder layout. An agent always belongs to a Data Product (LU); skills can also be shared across all LUs from Shared Objects.
Implementation/
LogicalUnits/<LU>/AI/
Agents/<agent>/
agent.yaml # the agent definition (opens in the Agent Editor)
Skills/ # skills private to this agent (*.html)
References/<pack>/ # reference packs searched at runtime (*.html)
Subagents/ # one <sub-agent>.yaml per folder sub-agent
Evals/ # eval suites (<suite>.csv)
Observation/ # signals.yaml and widgets.yaml for the Observation dashboard
Skills/ # LU-wide skills, available to every agent in the LU
References/ # LU-level reference folders
SharedObjects/AI/
Skills/ # skills shared by all LUs
The file formats themselves are documented in the Agent Platform section, starting with the agent.yaml Reference.
Open AI Explorer from its icon in the Activity Bar. The view has two roots, which are always shown so you can create the first asset even in an empty project:

Expanding an agent shows its folder as it is on disk: agent.yaml, Skills/, References/<pack>/, Subagents/, Evals/ and Observation/. Clicking a node opens it in the right editor:
agent.yaml and Subagents/*.yaml open in the Agent Editor.Evals/*.csv) opens in the Eval Editor. A suite node also expands into its test cases; clicking a case opens the Eval Editor on that case.Each test case under an eval suite carries a colored bullet that shows the result of its latest evaluation run. The color alone carries the meaning, so the tree stays easy to scan:
| Bullet | Meaning |
| Blue spinner | An evaluation is running. |
| Green | The last run passed. |
| Orange | The last run completed, but the case did not pass: one or more quality dimensions scored below their threshold, a conversation-level criterion was not met, an after hook failed, or the case was not evaluated (for example, because a before hook failed or the case has conversation-level evaluation without criteria). See Evaluation. |
| Red | The last run failed or was stopped. |
| Gray slashed circle | The test case is disabled. |
abc icon (no bullet) |
The case has not been evaluated yet. |

An agent whose agent.yaml references a missing or invalid sub-agent is decorated in red in the tree, and the problem is listed in the Problems panel. See Agent Editor.

Right-click a node in AI Explorer to see the actions that apply to it. For example, the context menu of an agent offers New Skill…, New Reference…, New Sub-Agent…, New Eval Suite…, Delete and Reveal in Explorer. Each New command asks for a name (letters, digits, underscore and hyphen; it must start with a letter), creates the file from a built-in template, opens it, and selects it in the tree.
| Command | Available on | What it does |
| New Agent... | Data Products root, an LU, an LU's Agents folder | Creates AI/Agents/<name>/ from the agent template and opens agent.yaml in the Agent Editor. Agents can only be created under an LU, never under Shared Objects. From the Data Products root you are asked to choose the LU. |
| New Skill... | An agent, a Skills folder, an LU, either root | Creates a skill .html file. On an agent or its Skills/ folder, the skill is private to the agent; on an LU it goes to <LU>/AI/Skills/; on Shared Objects it goes to SharedObjects/AI/Skills/. |
| New Reference... | An agent, a References folder or pack, an LU | Creates a reference .html file inside a reference pack. When invoked on a References root you choose an existing pack or New pack..., because a loose file directly under References/ is never loaded. References are not offered under Shared Objects. |
| New Sub-Agent... | An agent, its Subagents folder | Creates Subagents/<name>.yaml (same template as an agent) and opens it in the Agent Editor. |
| New Eval Suite... | An agent, its Evals folder | Creates Evals/<name>.csv with one empty test case, seeded with the agent's evaluation defaults, and opens it in the Eval Editor. The new case is evaluated at the conversation level and has no criteria yet: add conversation criteria (or switch the case to step level) before you run it. See Evaluation. |
| New Folder... | A References folder or pack | Creates a sub-folder, for example a new reference pack. |
| Delete | An agent, skill, reference file or folder, sub-agent | Deletes the item after a confirmation that names its kind (for example, "delete agent 'billing'"). Deleting a folder deletes all its contents. |
| Reveal in Explorer | Any node except the roots and test cases | Selects the backing file or folder in the standard file Explorer. |
| Refresh AI Explorer | The view toolbar | Reloads the tree. |
The New Agent, New Skill, New Reference, New Sub-Agent and New Eval Suite commands are also available in the standard file Explorer's context menu, on the matching folders: an LU's AI and AI/Agents folders, any Skills folder, References folders, an agent's Subagents folder and an agent's Evals folder.
Eval suites and test cases have their own context menus in AI Explorer:
New Test… creates a case evaluated at the conversation level, with no criteria. Add its criteria in the Eval Editor before you run it.


These commands edit the suite file and then soft-deploy the agent's LU automatically, so the next evaluation run sees the change. Test case authoring and running are covered in the Evaluation section.
New Agent… copies a template that gives you a working starting point:
agent.yaml with empty description, system_prompt, response_prompt and constraints, the built-in query and current_datetime tools, empty hook slots, planning enabled (without user approval), reference search enabled, and a settings: block. Comments in the template explain each key.Skills/, References/, Subagents/ and Evals/ folders.Observation/signals.yaml and Observation/widgets.yaml with generic samples. Replace them with your agent's own signals and widgets — the samples are used as-is otherwise. See Configuring Observation.The template's settings: block sets the following values, which differ from the runtime defaults that apply when a key is omitted (see Agent Editor — Settings):
settings:
llm_interface: default
snap_session_mode: none
memory: true
thinking: medium
use_cache: true
max_iterations: 30
max_retries: 2
logging_level: info
observability: true
max_step_continuations: 3
compaction_threshold_fraction: 0.8
compaction_context_window_tokens: 200000
compaction_keep_recent_messages: 30
compaction_interface: default
The agent runtime serves the deployed LU, not the files on disk. After you change agent.yaml, a sub-agent, a skill, a reference or a tool flow, deploy the agent's LU (a soft deploy is enough while iterating) before testing the change in Preview. Until then, Preview keeps answering with the previously deployed definition.
Eval runs are the exception: running an evaluation, and editing suites or cases from AI Explorer or the Eval Editor, soft-deploys the agent's LU for you so the run reads the latest suite files. The automatic soft deploy works only where the project sources are available (a Studio development space); elsewhere the run uses what is already deployed.
The following walkthrough shows how the pieces fit together when you build a new agent.
agent.yaml opens in the Agent Editor.query on for agents that answer from LU data, and use + Add flow tool to expose Broadway flows. For entity-scoped agents, set a step_start flow in Hooks (Before the user message) that loads the entity instance — its external parameters become the agent's inputs. See Hooks and Inputs.default). See LLM Connectors.If you prefer to start from a conversation with Studio AI, the agent-builder skill can perform steps 1–6 for you and then generate a first eval suite. See AI-Assisted Agent Development.
Next article: Agent Editor
AI Fusion Studio is the development environment for AI Fusion agents inside K2view Web Studio. Because an AI Fusion 3 agent is a set of files inside a Data Product (LU) — an agent.yaml, skills, reference documents, sub-agents, evaluation suites and observation configuration — you build it the same way you build the rest of your Fabric project: as files in your project tree. AI Fusion Studio adds the tooling that makes those files easy to create, edit, test and evaluate without leaving Studio.
Note: The examples and screenshots in this section use the banking demo (
customer_bank›banking). See About the Examples.
The tooling is delivered by the AI Fusion - Studio extension. It provides:
agent.yaml and sub-agent files. See Agent Editor.The extension is published on K2exchange as AI Fusion - Studio (version 1.0.2 at the time of writing) and requires Fabric 8.5.2 or later. Install it from the Extensions view of Web Studio, as described in K2exchange. The extension contributes no settings of its own — it talks to the Fabric server that hosts your Studio session.
Installing the extension also adds two Studio AI skills to your project, under .agents/skills/: agent-builder and researcher. They power the AI-assisted features described in AI-Assisted Agent Development. Commit them to Git together with the rest of the extension's files.
For the full list of AI Fusion prerequisites (the runtime extension, the LLM interface and the insights database), see Installation and Setup.
AI Explorer and the editors work on a fixed folder layout. An agent always belongs to a Data Product (LU); skills can also be shared across all LUs from Shared Objects.
Implementation/
LogicalUnits/<LU>/AI/
Agents/<agent>/
agent.yaml # the agent definition (opens in the Agent Editor)
Skills/ # skills private to this agent (*.html)
References/<pack>/ # reference packs searched at runtime (*.html)
Subagents/ # one <sub-agent>.yaml per folder sub-agent
Evals/ # eval suites (<suite>.csv)
Observation/ # signals.yaml and widgets.yaml for the Observation dashboard
Skills/ # LU-wide skills, available to every agent in the LU
References/ # LU-level reference folders
SharedObjects/AI/
Skills/ # skills shared by all LUs
The file formats themselves are documented in the Agent Platform section, starting with the agent.yaml Reference.
Open AI Explorer from its icon in the Activity Bar. The view has two roots, which are always shown so you can create the first asset even in an empty project:

Expanding an agent shows its folder as it is on disk: agent.yaml, Skills/, References/<pack>/, Subagents/, Evals/ and Observation/. Clicking a node opens it in the right editor:
agent.yaml and Subagents/*.yaml open in the Agent Editor.Evals/*.csv) opens in the Eval Editor. A suite node also expands into its test cases; clicking a case opens the Eval Editor on that case.Each test case under an eval suite carries a colored bullet that shows the result of its latest evaluation run. The color alone carries the meaning, so the tree stays easy to scan:
| Bullet | Meaning |
| Blue spinner | An evaluation is running. |
| Green | The last run passed. |
| Orange | The last run completed, but the case did not pass: one or more quality dimensions scored below their threshold, a conversation-level criterion was not met, an after hook failed, or the case was not evaluated (for example, because a before hook failed or the case has conversation-level evaluation without criteria). See Evaluation. |
| Red | The last run failed or was stopped. |
| Gray slashed circle | The test case is disabled. |
abc icon (no bullet) |
The case has not been evaluated yet. |

An agent whose agent.yaml references a missing or invalid sub-agent is decorated in red in the tree, and the problem is listed in the Problems panel. See Agent Editor.

Right-click a node in AI Explorer to see the actions that apply to it. For example, the context menu of an agent offers New Skill…, New Reference…, New Sub-Agent…, New Eval Suite…, Delete and Reveal in Explorer. Each New command asks for a name (letters, digits, underscore and hyphen; it must start with a letter), creates the file from a built-in template, opens it, and selects it in the tree.
| Command | Available on | What it does |
| New Agent... | Data Products root, an LU, an LU's Agents folder | Creates AI/Agents/<name>/ from the agent template and opens agent.yaml in the Agent Editor. Agents can only be created under an LU, never under Shared Objects. From the Data Products root you are asked to choose the LU. |
| New Skill... | An agent, a Skills folder, an LU, either root | Creates a skill .html file. On an agent or its Skills/ folder, the skill is private to the agent; on an LU it goes to <LU>/AI/Skills/; on Shared Objects it goes to SharedObjects/AI/Skills/. |
| New Reference... | An agent, a References folder or pack, an LU | Creates a reference .html file inside a reference pack. When invoked on a References root you choose an existing pack or New pack..., because a loose file directly under References/ is never loaded. References are not offered under Shared Objects. |
| New Sub-Agent... | An agent, its Subagents folder | Creates Subagents/<name>.yaml (same template as an agent) and opens it in the Agent Editor. |
| New Eval Suite... | An agent, its Evals folder | Creates Evals/<name>.csv with one empty test case, seeded with the agent's evaluation defaults, and opens it in the Eval Editor. The new case is evaluated at the conversation level and has no criteria yet: add conversation criteria (or switch the case to step level) before you run it. See Evaluation. |
| New Folder... | A References folder or pack | Creates a sub-folder, for example a new reference pack. |
| Delete | An agent, skill, reference file or folder, sub-agent | Deletes the item after a confirmation that names its kind (for example, "delete agent 'billing'"). Deleting a folder deletes all its contents. |
| Reveal in Explorer | Any node except the roots and test cases | Selects the backing file or folder in the standard file Explorer. |
| Refresh AI Explorer | The view toolbar | Reloads the tree. |
The New Agent, New Skill, New Reference, New Sub-Agent and New Eval Suite commands are also available in the standard file Explorer's context menu, on the matching folders: an LU's AI and AI/Agents folders, any Skills folder, References folders, an agent's Subagents folder and an agent's Evals folder.
Eval suites and test cases have their own context menus in AI Explorer:
New Test… creates a case evaluated at the conversation level, with no criteria. Add its criteria in the Eval Editor before you run it.


These commands edit the suite file and then soft-deploy the agent's LU automatically, so the next evaluation run sees the change. Test case authoring and running are covered in the Evaluation section.
New Agent… copies a template that gives you a working starting point:
agent.yaml with empty description, system_prompt, response_prompt and constraints, the built-in query and current_datetime tools, empty hook slots, planning enabled (without user approval), reference search enabled, and a settings: block. Comments in the template explain each key.Skills/, References/, Subagents/ and Evals/ folders.Observation/signals.yaml and Observation/widgets.yaml with generic samples. Replace them with your agent's own signals and widgets — the samples are used as-is otherwise. See Configuring Observation.The template's settings: block sets the following values, which differ from the runtime defaults that apply when a key is omitted (see Agent Editor — Settings):
settings:
llm_interface: default
snap_session_mode: none
memory: true
thinking: medium
use_cache: true
max_iterations: 30
max_retries: 2
logging_level: info
observability: true
max_step_continuations: 3
compaction_threshold_fraction: 0.8
compaction_context_window_tokens: 200000
compaction_keep_recent_messages: 30
compaction_interface: default
The agent runtime serves the deployed LU, not the files on disk. After you change agent.yaml, a sub-agent, a skill, a reference or a tool flow, deploy the agent's LU (a soft deploy is enough while iterating) before testing the change in Preview. Until then, Preview keeps answering with the previously deployed definition.
Eval runs are the exception: running an evaluation, and editing suites or cases from AI Explorer or the Eval Editor, soft-deploys the agent's LU for you so the run reads the latest suite files. The automatic soft deploy works only where the project sources are available (a Studio development space); elsewhere the run uses what is already deployed.
The following walkthrough shows how the pieces fit together when you build a new agent.
agent.yaml opens in the Agent Editor.query on for agents that answer from LU data, and use + Add flow tool to expose Broadway flows. For entity-scoped agents, set a step_start flow in Hooks (Before the user message) that loads the entity instance — its external parameters become the agent's inputs. See Hooks and Inputs.default). See LLM Connectors.If you prefer to start from a conversation with Studio AI, the agent-builder skill can perform steps 1–6 for you and then generate a first eval suite. See AI-Assisted Agent Development.
Next article: Agent Editor