Agents

An agent is an LLM with a system prompt, a model, generation parameters, declared input variables, and a set of allowed tools. Unlike a workflow, an agent does not follow a predetermined graph. The model decides which tools to call and in what order based on the conversation.

Agents live at /automation/agents. Each agent has a list view, a configuration/detail page, and a built-in test chat.

Use agents when the main challenge is understanding what the user wants, choosing the right tool, and explaining the result.

1

User asks

The user sends a message, or another system invokes the agent with context.

->

2

Agent chooses

The prompt, model, inputs, and chat history help the agent decide what to do next.

->

3

Tools run

The agent calls core tools or workflow tools to search, research, navigate, report, or perform a controlled action.

->

4

Answer returns

The agent explains what it found, what it changed, or what the user should do next.

Agent Anatomy

An agent's configuration covers the pieces that shape how it behaves:

Model-Specific Reasoning Controls

Reasoning controls change when you choose a different model because providers expose different options. The Agent Builder uses the selected model's capabilities to show only settings that apply:

• Claude Sonnet 5 uses adaptive thinking and a reasoning-effort setting. It does not use a token budget.
• Earlier Claude thinking models let you enable reasoning and set a token budget.
• OpenAI reasoning models let you choose reasoning effort and response verbosity.
• Google reasoning models use either a thinking level or a token budget, depending on the model.

When a supported model provides a reasoning summary during a streamed chat, the test chat displays that summary as it arrives. This helps you understand the model's approach without exposing its private chain of thought.

Agent Types

• Custom — built in your org for your use case.
• Suggested — recommended starting points.
• Customized System — a copy of a system agent you've modified.

Agents support tags and folder tags in the unified Automation Library, matching workflow organization.

Choosing Tools

The most important agent design choice is the tool set. Give the agent the smallest set of tools that can solve its job.

Tool selection rules of thumb:

• If the agent needs to answer or look something up, start with core tools.
• If the agent needs lightweight web discovery, use google_search; if it already has a known URL and needs the page text, use scrape_url. Use research_exa when the user wants synthesis with citations. When Genie relies on these external sources in its written response, citations appear as inline clickable source pills next to the relevant claims.
• If the agent needs to search, inspect, compare, or open the asset library, use content_library with a mode such as search, details, open_view, recommend_for_opportunity, compare, or find_duplicates. Set a purpose such as article_media, email_attachment, deal_enablement, deck_template, or internal_guidance so semantic search still enforces audience and allowed-use policy.
• If the agent needs reusable company-specific guidance, use skills. Skills are admin-managed markdown files that agents can find with search_skills and read with read_skill.
• If the agent needs to let the user browse asset results, use content_library with mode=open_view and the same purpose and filters to open a prefiltered library view. When the user names a folder, pass the folder path so the view opens scoped to that folder and its descendants.
• If the agent needs to personalize sales collateral, use content_library with mode=search and purpose=deck_template or purpose=deal_enablement, then use generate_deck when a PPTX deck template or fillable PDF should become a deal-specific generated asset.
• If the agent needs to create a new Content Studio image, use generate_image and insert its markdown with alt text or apply its URL to Open Graph metadata. generate_image can also use attached screenshots or reference images for edits and variations. If the agent needs an existing Content Library asset, reusable media/collateral, or approved image/video, use content_library with mode=search and purpose=article_media, then mode=details or fetch_library_asset to retrieve descriptions, tags, metadata, and embed-safe media URLs. If the user explicitly asks to save an attached file to the library, use upload_chat_attachment_to_content_library.
• If the agent needs to perform an operation, expose a workflow as a tool.
• If the operation has required business rules, keep those rules in the workflow, not only in the prompt.
• If the agent keeps calling the wrong tool, remove tools until only the useful choices remain.

Good Agent Jobs

Poor Agent Jobs

An agent is not the right primitive when the process should be identical every time.

Use a workflow instead when you need to:

• Apply the same transformation to every record in a batch.
• Enforce required field mappings or schema validation.
• Retry failed records with precise step-level history.
• Run from a trigger without a user conversation.
• Keep a process stable while you draft future changes.

Where Agents Run

Agents can be invoked from several places:

• Inside a workflow as an Agent task.
• From the test chat on the agent's configuration page.
• From Slack or Microsoft Teams via per-agent chat configuration in integration settings.
• From Genie surfaces in the app — the page-context Genie sidebar uses the same agent infrastructure under the hood.

Genie

Genie is the in-app AI assistant that appears as a floating button on authenticated pages. It is not literally one agent. It is a family of page-context agents that share the same core tool catalog. Each page can register its own Genie variant with a tailored system prompt, tool set, and structured page context, such as which account, report, or forecast slice the user is viewing.

On setup-heavy pages, Genie can show contextual suggestion bubbles next to the floating button. These bubbles are limited to once per user per day, and each user can snooze them for 7 days or turn them off from the bubble. Users can manage that setting from the Genie Tips panel.

Genie accepts typed prompts and short voice notes. Voice notes are transcribed into a reviewable prompt inside the composer before they are sent, so users can edit the text first instead of sending hidden audio.

The most common Genie surfaces include:

• Home Day Planner — generates a strictly prioritized plan for the user's day, drawing on profile, calendar, records, reports, and recent activity.
• Page Genies — forecast_page, account_page, opportunity_page, records_page, outreach_page, generate_accounts_page, etc. Each one knows about its page's data and can call page-aware tools, such as reading or updating contact org charts, inspecting files linked to account and opportunity records, explicitly uploading chat attachments to those records, creating or updating CRM records from record list pages when users ask, logging activities by resolving or creating the related contact/account, or filling UI controls with set_form_field.
• Record scratchpads — every account and contact has a Scratchpad tab: a free-form markdown notepad (tables and lists render) that your team can edit directly. On account and contact pages, ask Genie to save research, call notes, or plans there; Genie can read and update the scratchpad, and the tab refreshes after Genie saves.
• Workflow Builder Genie — code-defined agent embedded in the workflow builder. Lives in the codebase rather than the agents table.

Agent vs. Genie vs. Custom Agent

These three terms can blur together. The mental model:

• Genie is a label for the in-app, page-context AI experience. Under the hood it is implemented as a set of agents.
• Agents is the general term for any LLM + tools + prompt entity, whether it powers Genie, runs in Slack, or is a workflow step.
• Custom agent is what you call an agent you build at /automation/agents for your own use case.

Testing

The agent configuration page includes a chat interface that runs the agent against real org data and exposes per-message tool calls. Use it to verify:

• The agent understands the kinds of questions users will ask.
• It reaches for the tools you expect.
• It avoids tools it should not need.
• It asks for missing inputs instead of guessing.
• It explains actions clearly after a workflow tool runs.

Next Steps

• Workflows vs. agents — when to use which.
• Workflows — how to build controlled actions agents can call.
• Build your first agent — guided walkthrough.
• Build a multi-tool agent — combine core tools, workflow tools, and trace inspection.
• Use a workflow as an agent tool — compose a conversational agent with a workflow action.