AI Agent

Give an autonomous AI Agent a goal and tools so it can reason and act. It reads your instructions, calls tools, and stops when done or when a limit is hit. For one generated answer without tools, use AI Prompt.

At a glanceVerified in app · Oct 29, 2026

Use it to

Pursue a goal across multiple turns using web search, app actions, subflows, MCP tools, and memory

Find it

Tools → AI → AI Agent

Cost

0 Step credits for the Agent Step · Uses Glow AI tokens (plus Step credits for paid app tools)

Limits

Defaults: 10 iterations (max 50) · 40 model calls (max 200) · 25 tool calls (max 200) · 180 s active time (max 600 s)

Output

{{ N.result }} (or structured schema fields), {{ N.estimatedCreditsUsed }}

Instead

AI Prompt (single-turn generation, classification, or extraction without tools)

Attribute	Value
Default model	Gemini 4 Argon, Glow managed (or connect OpenRouter)
Runs per item	No. Place inside a Repeater
Typical cost	~14 credits without tools; 150 to 460 with search or app tools (Cost)

Setting it up

Configure the agent across four tabs in the App drawer:

1. Instructions: Write the goal, input (Classify this email: {{ item }}), and completion criteria. Identity, Style, and other guideline chips are optional.
2. Tools: Pick the helpers, app actions, workflows, or MCP tools the agent may call.
3. Context: Enable Conversational Memory or attach reference files.
4. Output: Choose Text, JSON, or Separate fields to validate the reply shape.

Tools and approvals

Tool type	Included capabilities	Default
Built-in helpers	Search the web, Read a web page, Run code, Search reference files, Recall conversation	On (turn off unused ones)
App actions	Actions from connected apps like Gmail or Slack	Off (add via Add tool)
Glow workflows	Reusable automations called as a Subflow	Off
MCP tools	Tools from an external MCP server registered by an Admin	Off

Set approvals per account and action in Tools, never in the prompt. Read-only actions (finding an email, listing Slack users) run immediately. Write actions with approval enabled pause the run until someone clicks Approve or Reject in Executions or Approvals.

What it passes on

The agent returns its answer in {{ N.result }} and its credit estimate in {{ N.estimatedCreditsUsed }}. Inside a Repeater, When finished lists every pass under {{ N.results }}. Intermediate reasoning stays internal.

Answer format	Behavior	Downstream reference
Text	Plain prose reply without JSON parsing.	{{ 2.result }}
JSON	Parses JSON without enforcing specific fields.	{{ 2.result.field }}
Separate fields	Validates the reply against declared fields and types.	{{ 2.result.lead_score }}

With Separate fields selected, define the schema in one of three ways:

• Fields: Add rows, pick a type (text, number, yes/no, list, group), and mark required fields.
• Advanced (JSON Schema): Paste or edit a JSON Schema object.
• From a previous run: Click Use this to prefill from the last output, or Set it up myself.

Mark fields required before using them in a Condition. Switching to Text or JSON keeps your saved schema inactive until you re-select Separate fields (optional shape repair).

Limits

Adjust bounds under Advanced runtime limits (clear a field to restore its default):

Limit	Default	Range	Behavior
Iterations	10	1 to 50	Caps reasoning turns per run.
Model calls	40	1 to 200	Counts main turns, helper sub-agents, retries, fallback, and output repair.
Tool calls	25	0 to 200	Counts proposed, denied, and corrected calls (resuming an approval counts 1).
Active time	180 s	10 to 600 s (excludes approval wait)	Caps active runtime and reserves the final 20% (6 to 25 s) for the reply.
Estimated credits	Off	1 to 1,000,000	Stops new turns once estimated usage hits the cap (in-flight calls finish).

Hitting any cap fails the Step with AGENT_LIMIT_REACHED, saves partial work in partialOutput, and skips automatic retries.

Cost

The Step header estimate (About 14 AI-derived shared credits / run) assumes no tools. Equipped tools add schemas and results to each turn:

Scenario	Typical credits	Why
Classify a short email, no tools	14 to 15	One model call with prompt and reply tokens.
Same task with web search	~162	Runs a search model call plus a synthesis turn.
Same task with the Slack tool (list users, DM)	~458	Sends app tool schemas and lookup outputs on each turn.

Check {{ N.estimatedCreditsUsed }} or Executions after a run. Disable unused helpers, pick Selected actions instead of Whole connection, and set an Estimated credits cap. Review spend in Billing & Usage.

---

The AI Agent canvas card

Click any tile on the card to open its section in the App drawer:

• Instructions: Click Add → or Edit → to edit the brief.
• Tools: Shows {N} apps · {M} total actions across apps, subflows, and MCP tools.
• Memory & Reference sources: Opens Context for memory and workspace files.
• Answer format: Opens Output to pick Text, JSON, or Separate fields.
• Header & Footer: Header holds model selection, Executions, Test & Debug, and ▶ Run. Footer has Run test → and Inspect → (Test an AI Agent).

Shape the behaviour

In Instructions, click any + Label chip under Advanced prompt guidelines:

Leave Conversational Memory off unless runs share history (choose which runs share memory).

---

Execution safeguards and recovery

Duplicate action safeguards

A new run does not undo earlier tool calls and can repeat writes. Check Executions before retrying (recovering from a failed run).

Automatic output schema validation & repair

Separate fields validates the reply against your schema, while JSON only parses it. Optional repair runs at most one extra AI call without re-running tools.

Can the AI Agent run for each item?

AI Agents run once per execution and do not use Run for each item. Place the agent inside a Repeater or call a Subflow:

---

How the agent reaches your accounts

App tools use the connection bound to each tool. Glow verifies permissions before every call and rejects disconnected or removed accounts.

Looking up values for an action

Activity entries labeled Look up values: … are read-only option lookups (folder, label, or sheet IDs). If a lookup finds no match or needs a parent field, paste the exact ID in Instructions.

Approving what the agent does

Use Human Review to gate downstream Steps. Never equip an unapproved tool for an action that a downstream Human Review Step gates.

Examples to copy

Copy prompts and Separate fields schemas from the testing guide:

Verify it works

Run

Replace the default instruction, click Run test on the card, and choose Preview

You should see

Read-only tools run and return {{ N.result }} and {{ N.estimatedCreditsUsed }}, while writes stay Proposed.

What’s Next?

• 👉 Test an AI Agent →: Preview, Live test, and approvals.
• How testing works: compare Testing data, Pin output for tests, Preview, and Run.
• Credits & Allowances: see how Step credits and Glow AI tokens are billed.