Deploying your first agent

This guide takes you from choosing an agent → running it reliably → shipping it into a real workflow.

Browse agents Use cases

Step 1 — Pick the right agent

Use these questions to avoid picking an agent that looks good but fails in production:

• What input will you have every time (text, URL, JSON)?
• What output do you need (plain text, structured fields, citations)?
• What’s “good enough” accuracy for the workflow?

Pro tip: start with a narrow task. Expand after you have one stable success case.

Step 2 — Run it manually (fast iteration)

1. Open an agent in Search
2. Run it with a real example from your workflow
3. Save:
   • the exact input you used
   • the output you liked
   • the output you didn’t like

Checklist

• Save 3 successful examples
• Save 3 failure examples
• Write down the acceptance criteria (what makes output usable)

Step 3 — Add guardrails to your inputs

Before you automate, make your inputs predictable:

• Strip irrelevant text
• Normalize formatting (dates, currencies)
• Include a clear objective (“extract fields”, “draft reply”, “classify intent”)

Example JSON input pattern:

{
  "objective": "Draft a friendly support reply",
  "customer_message": "...",
  "company_policy_excerpt": "...",
  "tone": "concise, warm, accurate",
  "constraints": ["don’t promise refunds", "ask one clarifying question if needed"]
}

Step 4 — Ship it into a workflow

Pick one integration path:

• Human-in-the-loop: agent drafts, human approves
• Automation: agent output triggers actions (only after you’re confident)

Integrate via API

Recommended agent to start with

Common pitfalls

Pitfall	Why it happens	Fix
Output is inconsistent	Input varies too much	Normalize inputs; add examples
It hallucinates details	Missing context + no constraints	Provide policy/context; add constraints
It’s expensive	Too many tokens	Shorten context; cache; chunk requests

Next steps

• Customizing agents
• Security overview