Gen3 Agents Building

Building Agents

Start Here

1. Open Agents → sidebar Configuration (or Agent Configuration view).
2. Select Create Agent and choose Agent Configs for a blank form or Agent Templates to preload from the catalog.
3. Set instructions, model, datasets, category (create, edit, or delete inline when allowed), modality tabs, and Web Search Settings as needed.
4. Save the agent and test immediately in GT Chat.

Why this matters

Well-built agents encode repeatable behavior so operators do not re-type complex prompts and policy in every conversation. Gen 3 separates quick template preloads from deep capability tuning so you can start fast without guessing which tools and models apply.

Details

Use the Agent Configuration workspace on Agents when you need to create a new agent or revise an existing one. In the active Gen 3 product, agent creation is a practical workflow: pick a starting point (blank or template), define behavior, choose the right model posture, attach datasets, enable managed tools where appropriate, then validate the result in chat.

Start from a template

When you select Create Agent, the configuration panel exposes two top-level tabs:

Tab	Use when
Agent Configs	You want a blank form or you already applied a template and are editing the preloaded fields.
Agent Templates	You want to browse the live template catalog and preload an agent from a CSV starter.

Catalog source and refresh

Templates are not bundled inside the tenant app. The catalog is fetched live from GT-Edge-AI/GT-AI-OS/agent-templates/*.csv through the tenant API (Control Panel proxies GitHub on behalf of the tenant). Opening Agent Templates triggers a fresh catalog fetch; use Refresh templates to retry after network or GitHub errors.

New .csv files added to agent-templates/ in the instructions repository appear automatically on the next catalog load—no tenant redeploy required.

Preload matrix (CSV → editor)

The tenant API reads the header row and first data row of each template CSV. Mapped columns populate the agent editor:

CSV column(s)	Editor field
name	Agent name
description	Description
category	Category (or create inline in the agent editor via + Create new category…)
tags	Tags (comma- or pipe-delimited)
model / model_id	Chat model (resolved against tenant chat models by id, model id, or display name)
prompt_template / system_prompt / systemprompt	System prompt
disclaimer	Disclaimer
temperature	Temperature
easy_prompts / easy_buttons	Easy buttons (pipe-delimited)
visibility	Visibility (individual/private → private; team/group → group; organization/tenant → tenant)
selected_dataset_ids / dataset_ids / selected_dataset_names / dataset_names	Attached datasets (UUID or dataset name match)
web_search_access / web_search_agent / web_search_enabled	Web search access (enabled / disabled). When omitted, matches the Control Panel Enable web search tool for agents by default setting (same as a blank new agent).
web_search_model_id / web_search_model	Optional override for the agent web search model (resolved like model)
default_memory_mode / memory_mode	Default Contextual Memory scope for new chats (this_conversation, this_agent, all_agents; legacy memory_type values are normalized). When the mode implies recall, Enable Contextual Memory for this agent is turned on after apply.
contextual_memory_allowed	Optional explicit Enable Contextual Memory for this agent (true / false). When set, overrides inference from default_memory_mode.
vision_model_id / vision_model	Agent Vision Settings model (resolved against tenant vision models)

Ignored columns: max_tokens, dataset_connection, and category_description are present in legacy template CSVs for compatibility but are not applied to Gen 3 agent records.

Capability templates: Starters that need managed tools set explicit CSV flags where possible (for example internet-search-agent.csv sets web_search_access=enabled). Speech, image generation, and vision still require a configured model on the matching modality tab when the deployment does not supply a usable default—vision-chat-agent.csv reminds operators to pick a vision model after apply.

Selecting Use template switches you to Agent Configs with a Template applied banner so you can review, adjust, and save.

Available templates (8)

Template file	Purpose
document-chat-agent.csv	Document Q&A with RAG over uploaded files
internet-search-agent.csv	Web research with sourced answers (enable web search separately after apply)
system-prompt-reviewer.csv	Review and improve system prompts
nemotron-agent.csv	Local Nemotron on Ollama for advanced reasoning
nemotron-mini-agent.csv	Fast Nemotron Mini on Ollama
python_coding_microproject.csv	Python / Streamlit coding tutor
vision-chat-agent.csv	Vision-enabled RAG over image datasets (configure vision tab after apply)
kali_linux_shell_simulator.csv	Simulated Kali Linux terminal for training

Template failure modes

Symptom	Likely cause
Catalog fails to load	Tenant cannot reach Control Panel / GitHub; check network and retry Refresh templates.
Empty catalog	No .csv files in agent-templates/ or GitHub listing error.
Card shows summary error	Individual CSV could not be parsed (malformed quoting, missing data row).
Apply fails	CSV missing required header/data row or tenant API rejected the fetch.
Model field empty or wrong	Template model string does not match any tenant chat model; pick a model manually on Chat Model Settings.
Datasets not attached	Names or UUIDs in CSV do not match datasets in your tenant.

Templates preload text and defaults only—they do not create datasets or publish models.

Agent capabilities and tools

Inside Agent Configs, Agent Settings uses modality tabs. Model pickers label deployment defaults as Default Vision Model, Default STT Model, Default TTS Model, Default Image Generation Model, and Default Web Search Model so you can distinguish platform baselines from explicit per-agent overrides.

• Chat Model Settings — chat model, temperature, system prompt
• STT / TTS Settings — speech input and output models
• Vision Settings — vision model for image understanding
• Image Generation Settings — image creation defaults
• Web Search Settings — managed public-web research tool
• Contextual Memory (when the platform layer is enabled) — per-agent allow gate, default-on for new chats, and default recall scope

Contextual Memory (agent configuration)

When your deployment's Contextual Memory platform layer is enabled (Control Panel Super Admins configure indexing, embedding models, and TTL), Agent Settings exposes a Contextual Memory section:

Control	Purpose
Enable Contextual Memory for this agent	Master allow for this agent. When off, the GT Chat brain control is hidden and memory tools cannot run for this agent—even if the platform layer is on.
On by default for new chats	When on, new conversations start with memory enabled; when off, users opt in per chat from the brain control.
Default recall scope	This conversation only, Just conversations with this agent, or All my conversations when memory is on by default.

Users can still override per conversation from the brain control when memory is allowed for the agent.

Memory modes and agent identity

Contextual Memory has three layers: platform (Control Panel indexing), agent (this section), and per-conversation (brain control in GT Chat). The active mode for a chat determines what prior threads can be recalled:

Per-chat scope (brain control)	What gets recalled	Agent identity note
Memory off (current chat only)	Only messages in the open thread	No cross-thread recall; the chat agent uses this conversation's history only.
Just conversations with this agent	Prior chats the user had with this same agent	Assistant lines in recalled excerpts are that agent's prior replies.
All my conversations	Prior chats across every agent the user has used	Excerpts cite agent name • conversation title • date. A chat agent must not treat assistant replies from another agent's citation as its own prior behavior—the cited agent name identifies who the user was talking to.

GT Chat injects ACTIVE CHAT CONTEXT at runtime so the model knows which agent it is and which memory scope is enabled. Operators configuring agents should match default scope to the job: use Just conversations with this agent for specialized assistants; use All my conversations only when cross-agent continuity is worth the attribution risk.

GT Helper and tenant chat agents are separate surfaces—memory about product coaching in GT Helper is not the same as memory from a tenant chat agent unless the citation says so.

Managed chat tools (tenant UI)

Gen 3 exposes a fixed set of managed tools the chat model may call during GT Chat. The agent configuration UI does not expose arbitrary MCP integration pickers—operators tune behavior through datasets, modality tabs, and web search gates below.

Tool	When the model uses it	Operator notes
search_datasets	Retrieve and count evidence from conversation-attached datasets, agent default datasets, or staged upload documents in the current chat scope.	Requires dataset or document scope in the thread. Not a substitute for public-web research.
analyze_image	Direct vision review of images already stored in scoped datasets.	Requires vision model configuration and image-bearing documents.
generate_image	Create a new image from a text prompt using agent or deployment image-generation defaults.	Requires image generation model configuration.
web_search	Live public-web research through the deployment or agent web search model.	Requires deployment web search model configuration and agent Web search access set to Enabled (below). Separate from dataset retrieval.

No MCP executable tools in the agent builder: Unlike legacy flows that surfaced configurable MCP integration ids in the UI, Gen 3 tenant agent configuration does not let operators attach arbitrary MCP servers. Managed tools above are executed by the platform when policy and model capabilities allow.

Web search access

The managed web_search tool runs only when:

1. Deployment model — Control Panel Models → Default Models selects a web search model (and may enable web search for new agents by default; see Models).
2. Agent access — Web Search → Web search access is Enabled for that agent (the UI exposes only Enabled and Disabled—no inherit mode). New agents start Enabled or Disabled according to the Control Panel Enable web search tool for agents by default setting.

If deployment web search is off or the agent access is Disabled, chat will not invoke web_search for that agent even when the user asks for current events.

Chat tool timeline

During streaming replies, GT Chat surfaces tool activity on the assistant message (for example “Search datasets”, “Web search”, “Generate image”). Completed, skipped, and failed tool steps appear in the activity strip so you can see whether retrieval ran, was deduplicated, or failed before the final answer.

Core build sequence (blank form)

1. Give the agent a clear name that reflects the work it should do.
2. Write instructions that define the job, constraints, and expected style.
3. Choose the appropriate model or runtime configuration.
4. Attach only the datasets that should travel with the agent by default.
5. Configure modality and web search tabs when the job needs vision, speech, images, or public-web research.
6. Save the agent.
7. Test it in GT Chat.

What makes a good Gen 3 agent

Specific instructions

Agent instructions should describe a concrete operating posture, not a vague personality. Tell the agent what sources it should rely on, what kinds of outputs it should produce, and any workflow constraints that matter.

Intentional dataset defaults

Attach default datasets only when they should be part of nearly every use of the agent. If the source material changes per conversation, leave the agent lean and attach datasets from chat instead.

Model fit

Use the model that matches the task. The goal is reliable execution, not maximum size by default.

Build-review loop

After every meaningful change:

1. launch the agent in GT Chat
2. ask a prompt that represents the real use case
3. confirm the behavior, tool usage, and source grounding are correct
4. return to configuration only if the result shows a real gap

Common mistakes

• attaching too many datasets to one agent
• writing instructions that are too broad to produce repeatable behavior
• expecting web research from search_datasets instead of enabling web_search
• changing multiple variables at once and then not knowing which change helped
• forgetting to re-test after model, dataset, or tool-setting changes

Best practices

• Build separate agents for separate jobs.
• Keep default retrieval context small and purposeful.
• Use favorites only after the agent is stable enough for regular use.
• Export important agents before large refactors.
• After applying a template, walk every modality tab before publishing to other users.

Related pages

• Agents
• Sharing Agents
• Using Datasets in Chat
• Control Panel Models
• GT Helper