Agent Forge

Super Agent

Catch-all Super Agent with Lite Mode tool access inside workflows

The Super Agent block (type ask_agent) brings Lite Mode ask-agent capabilities into Workflow Mode. It uses the same copilot server tools as Lite chat — web search, documentation search, Google Workspace, Slack, Notion, Jira, Firecrawl, and more — without manually wiring each integration block.

Super Agent uses copilot/lite billing (per agent call and per tool invocation), not standard workflow provider execution billing. Costs appear in workflow execution logs like other copilot usage.

Overview

Use Super Agent when you want a single workflow step to reason, search, and act across many integrations:

Catch-all automation: One block handles research, lookups, and light integrations

Implicit toolkit: No tool picker — the full Lite ask-mode server toolkit is available automatically

Skills support: Attach Skills from Settings, same as Lite Mode and the Agent block

Safety toggle: Read-only by default; enable mutations for write operations

Observable output: Returns content, tool calls, tokens, and model used

Unlike the workflow Copilot panel (which builds and edits workflows with mode: 'agent'), Super Agent runs a headless ask loop designed for production workflow execution.

Configuration

System Prompt

Optional instructions that shape how the agent behaves on every run. Supports the prompt wand for AI-assisted drafting, same as the Agent block.

User Prompt

The task or question for this step. Required. Supports references to upstream block outputs, for example <starter.input> or <router1.content>.

Skills

Optional Skills from Settings. Their markdown is loaded at runtime and appended to the system prompt.

Model

Override the default Lite-style model. Supported models include:

  • gpt-5.4
  • claude-opus-4-8
  • gpt-5.1
  • claude-sonnet-4-5-20250929

When left at the default, Depth selects the model via copilot depth configuration.

Depth

Depth level 03 maps to copilot depth models when no explicit model override is set:

DepthLabel
0Fast
1Standard
2Deep
3Maximum

Allow Mutations

When off (default), mutating operations are blocked before execution — send email, update sheets, post to Slack, non-GET API calls, and similar writes. Read-only tools still run.

When on, write operations execute automatically. Use this for production automations that need to send messages, update records, or call mutating APIs.

set_environment_variables always requires Allow Mutations to be enabled, even when other read tools are allowed.

Max Tool Rounds

Safety cap on how many tool calls the agent may make in one run. Default: 20. Combined with a 120 second execution timeout, this prevents runaway tool loops from blocking your workflow.

Memories (Advanced)

Optional connection to Memory block output for additional context in the user/system message pipeline.

Available Tools

Super Agent exposes the Lite Mode ask server toolkit. Categories include:

Search and research

  • search_online, search_documentation
  • arxiv, wikipedia, youtube, firecrawl

Utilities

  • make_api_request
  • get_environment_variables, set_environment_variables (mutations required)
  • get_oauth_credentials

Google Workspace

  • google_gmail, google_drive, google_docs, google_sheets, google_calendar
  • list_gdrive_files, read_gdrive_file

Integrations

  • slack, notion, jira, confluence, airtable, dropbox
  • hubspot, linkedin, wordpress, x

Media and parsing

  • image_generator, elevenlabs, mistral_parse

Markets

  • polymarket (prediction market data)

OAuth-backed tools require the corresponding integration to be connected in workspace Settings, same as Lite Mode.

Not Available in Super Agent (v1)

These tools are excluded from workflow execution:

ToolReason
create_artifact, edit_artifactBrowser artifact panel only
build_workflow, edit_workflow, get_user_workflowWorkflow copilot only
run_workflowRecursive execution risk
gdrive_request_accessOAuth UI flow

Browser-only or client-side tools return a structured error in toolCalls rather than failing silently.

Outputs

OutputTypeDescription
contentstringFinal assistant text response
modelstringModel used for the run
tokensjsonprompt, completion, and total token counts
toolCallsjson{ list, count } — each entry includes name, success, result, and error when applicable

Reference outputs in downstream blocks using your block's canvas ID, for example <superagent1.content>.

Super Agent vs Agent Block

AgentSuper Agent
Tool selectionYou pick specific tools via the tool inputFull Lite ask-mode server toolkit (implicit)
RuntimeDirect LLM provider requestAgent ask loop with server tool routing
BillingWorkflow / provider executionCopilot / lite billing
API keysPer-provider keys on the blockUses workspace copilot + OAuth connections
Best forPrecise, repeatable tool useOpen-ended catch-all tasks

Example Workflow

Connect Starter → Super Agent → Response:

  1. User prompt: Research the latest news on {topic} and summarize in 3 bullet points
  2. Allow Mutations: off (read-only search)
  3. Response block body: <superagent1.content>

The agent searches the web and returns a summary without wiring Serper, Firecrawl, and a separate Agent block.

Automation with Writes

For a workflow that posts to Slack after research:

  1. Enable Allow Mutations
  2. Prompt: Summarize today's news on AI and post the summary to #general
  3. Ensure Slack OAuth is connected in Settings

Billing and Limits

  • Per agent call: Copilot billing with referenceType: ask_agent_block
  • Per tool call: Copilot tool billing on each successful server tool execution
  • Usage gate: Copilot usage allowance is checked before execution (workflow runs do not use Lite Mode's per-minute chat rate limit)
  • Timeout: 120 seconds per block execution

Review token and tool usage in workflow execution logs.

Troubleshooting

Tool blocked with mutations disabled

Enable Allow Mutations or rephrase the task to use read-only operations (search, get, list).

OAuth tool fails

Connect the integration in workspace Settings. Errors surface in toolCalls and may appear in content.

Hit max tool rounds

Increase Max Tool Rounds or simplify the prompt so the agent completes in fewer steps.

Cost higher than expected

Each tool invocation is billed separately. Narrow the prompt or use the standard Agent block with explicit tools for predictable cost.

Notes

Super Agent