Describe the agent you want and let AI build it, or design it yourself on a visual canvas — the model, instructions, tools, output and the rules every answer must pass. Test it, see that it's ready, then let it run on triggers and schedules, call your workflows and APIs, and answer plain-English questions about your data.
Design, test and ship AI agents
Reads supplier invoices from email and drafts bills.
Hands new orders to the stock, billing and approval agents.
Scores every new lead the moment it's created.
Answers order questions on WhatsApp. Never promises refunds.
with attachment
guardrails on
6 fields · line items
2 read · 1 write
totals must add up
An agent is a saved configuration, not a black box. You pick the language model, write the system instructions that shape its behavior, and grant it a precise set of tools — your workflows, queries and connectors. Nothing more. It only ever sees the entities and rows your permissions allow. Use Simple mode for the essentials, or Advanced to see every setting.
Anthropic Claude, OpenAI, Google Gemini or AWS Bedrock, using your own API keys from the secrets manager. Set temperature and output limits, and write the instructions and guardrails the agent must follow.
Reasoning agents loop through their tools until the job is done. Reflex agents run their tools once and answer — fast and predictable. Orchestrators plan across your other agents and tools, then write one answer.
Attach workflows, saved queries and connector actions. Each is labelled read only, destructive or safe to repeat, and the agent can only call what you've explicitly granted.
Conversations carry context across turns, and every action respects your RBAC and row-level access.
Say what the agent should do in plain English. The builder drafts the whole configuration — model, instructions, tools, output and rules — and checks it before you see it. Ask for changes the same way.
See every change first. Edits arrive as a diff you apply or discard — nothing is saved until you press Save.
Asks instead of guessing. It asks clarifying questions when your request is vague, and lists the assumptions it made.
Flags anything that writes. You're warned whenever an agent can change data through one of its tools.
Or start from a template. Read documents into data, coordinate other agents, or run one quick pass — then fine-tune by hand.
Rules catch bad answers before they reach your data, the test panel shows you exactly how the agent behaves, and a readiness score tells you what's still missing.
Rules check each answer against your output shape. When one fails, the model gets your retry message and tries again. Rules are suggested for you from the output schema.
Run the agent with a message, parameters and file attachments, keep the conversation going, and get a pass/fail scorecard against every rule.
Live checks flag errors and warnings as you build, and a readiness score shows what's done and what's left before you rely on it.
Agents follow an agentic loop: understand the request, decide which tool to call, run it against your data, then reason over the result to answer or act.
Agents translate plain English into queries over your entities, then render the result the way a human would want to see it. No SQL, no report builder, no waiting on analysts.
Grounded on your model. Questions resolve against your real entities, fields and relationships — answers are specific, not hallucinated.
Tables & charts on demand. "Break revenue down by region" returns a chart; "list overdue invoices" returns a table — automatically.
Follow-ups keep context. "Now just the ones over $20k" refines the previous answer without repeating yourself.
Respects access. Each user only ever queries the rows and fields their permissions and hierarchy allow.
Because agents call your real workflows and connectors, they can take action on the result, end to end.
Send a WhatsApp template, an email, or a Slack message — fired straight from a workflow the agent triggers.
Create or edit records, change a status, or move an item through your process — every write checked against permissions.
Hit REST APIs, charge through Stripe, read from entities — using secrets from the secrets manager.
Wake every morning to chase overdue invoices, summarize yesterday's tickets, or post a daily pipeline digest to Slack.
Trigger on create, edit or status change — a new lead, a stuck ticket, a high-risk order — and act within seconds.
Hold a real conversation in the inbox — text, images, audio, documents — keeping context across the whole thread.
Runs daily, finds invoices overdue more than 30 days, and sends each customer a WhatsApp reminder with the amount and due date.
On every new ticket, classifies priority, suggests a category, and routes high-severity issues to the on-call queue.
Answers "which deals slipped this month?" with a table, then drafts follow-up emails for the reps who own them.
The two-way inbox where agents hold threaded conversations with text, media and documents.
Explore →Event triggers, schedules and custom actions that wake agents and fire on their output.
Explore →The pivots and saved queries that agents call as tools to read your data.
Explore →The connectors and API calls — wrapped in workflow steps — that agents invoke to take action.
Explore →