Agents that work your way.
Describe how the work should go, and your AI agent follows it, with your data, your systems, your emails and the right person’s sign-off, every step on record.
No code, and no new AI to buy. Works with Claude Desktop, Codex, Cursor, OpenClaw and 15 more. Free during the pilot.
Run by your AI agent
Refund request · order #5521
Your AI agent
Read the order
Looking up order #5521…
Jev · decision AI
Assess the claim
Next
Your AI agent
Apply your refund policy
Next
Priya · Finance
Approve the refund
Next
Your AI agent
Issue the refund
Next
Your AI agent
Tell the customer
Next
Illustrative example
What stops teams handing work to agents
- Without a processIllustration
Done! Refund issued.
Issue the refund · required to finish
- Refund receiptMissing
- Amount matches the orderNot checked
Sent back to the agent
Done! Refund issued.
Issue the refund · required to finish
- Refund receiptMissing
- Amount matches the orderNot checked
Sent back to the agent
How a refund gets done
First, tell your agent how the work goes.
You, to your agent
When a refund request comes in, read the order and have Jev check the claim. Anything over $500 goes to Priya in Finance. Then issue the refund and tell the customer.
Refund requestDraft process
- Read the orderYour AI agent
- Assess the claimJev · decision AI
- Over $500 goes to FinanceLabelled route
- Approve the refundPriya · Finance
- Issue the refundYour AI agent
- Tell the customerYour AI agent
Then connect what it needs.
You, to your agent
Read orders from our orders database. Use the payments API to issue refunds, and our refund template to tell the customer.
What this process may use
- OrdersPostgreSQLRead
- Payments APIRESTIssue refund, in this process only
- Refund confirmationYour email templateSend to the customer
Passwords and keys stay with Agent Process. The agent never sees them.
Then let it run.
You, to your agent
Handle refund request #5521.
CompletedIllustrative run record
Refund request · order #5521
Your AI agent · orders database
Order read
Order #5521 · $740.00 · arrived damaged
Jev · decision AI
Claim assessed
Damaged in transit · 94% sure · above your 70% line
Your AI agent
Route chosen
Over $500, so the route to Finance; reason recorded
Priya · Finance
Approved, from email
Decision + what Priya saw, kept together
Your AI agent · payments system
Refund issued
Receipt re_8f2
Your AI agent · your template
Customer told
“Your refund is on its way” · send recorded
Open any run to see what the agent did, the data it used, how Jev judged it, the route it took, who approved and which emails went out. Each organisation’s data is kept apart, and every approval keeps what the person saw.
Security and boundariesWhat the pilot supports today
- Agents connect through MCP with an access token. ChatGPT and claude.ai connectors and scheduled starts are not enabled in the pilot.
- Controls apply to work routed through Agent Process. We do not control an agent’s other tools or guarantee the truth of its reports.
- Database sources support reads. Enabled REST endpoints support writes. Published email templates support outgoing communication; handling replies is not supported yet.
Priya approves it from her inbox.
The refund is over $500, so the run waits for Finance. Priya gets an email, opens the request with the order and Jev’s assessment in front of her, and decides. She needs no account. Try it as Priya.
Priya’s review inbox
Refund process / Order #5521
Refund $740.00 for order #5521
The customer reports the order arrived damaged. Over $500, so Finance decides.
Agent recommendation
Issue the full refund.
Jev assessed the claim as damaged in transit, 94% sure, above your 70% line.
- Order
- #5521 · $740.00 · delivered 2 May
- Jev
- Damaged in transit · 94% sure
- Route
- Over $500 goes to Finance
Assigned to Priya
Finance approval · refund request
You’ll confirm your choice in the next step.
Run record · order #5521
- Waiting for approval Assigned to Priya
- Issue the refund Waits for the decision
Waiting for you. Opening a request doesn’t decide it.
Jev makes the judgement calls
- Sort
- “Which kind of request is this?” Picks one of your categories, so the process takes the right path.
- Score
- “How well does this meet our standard?” Rates work on your own scale, from your own descriptions.
- Check
- “Is this true, given the evidence?” Says how likely a statement is, so weak cases don’t slip through.
When Jev is unsure, a person decides, so unclear never passes on its own. You set the lines, and your process decides what each answer means. Agents can’t submit or overwrite its answers, and each one is kept with the run.
Jev · decision AI
Your question: what is this request about?
Refund request 1 of 4
“The vase arrived in pieces and the box was crushed.”
- Damaged in transit96%
- Changed their mind3%
- Never arrived1%
- Your line: 70% sure
Damaged in transit
Refund path
0
Changed their mind
Returns policy
0
Unsure
A person decides
0
Works with the agents you use
How to connect each one- OpenClaw
- Hermes Agent
- NanoClaw
- nanobot
- ZeroClaw
- Pi
- Claude Code
- Codex
- Gemini CLI
- OpenCode
- Goose
- Claude Desktop
- Cursor
- VS Code
- Cline
- Zed
- Devin Desktop
- LM Studio
- TypeScript SDK
No new AI to buy. Agents connect through MCP, the open standard for plugging AI agents into business tools, and setup takes a few minutes. You can also give a client’s or partner’s agent a key to the processes you choose, and pause or revoke it at any time.
Questions
Agent Process turns your business processes into steps the AI agents you already use can follow. Your agent does the work with your data, systems and email templates, Jev makes the judgement calls against your criteria, and the right person signs off before anything important. Every step is recorded.
What is Agent Process?
Agent Process turns your business processes into steps the AI agents you already use can follow. Your agent does the work with your data, systems and email templates, Jev makes the judgement calls against your criteria, and the right person signs off before anything important. Every step is recorded.
19 agents, including Claude Desktop, Claude Code, Codex, Cursor, OpenClaw, VS Code and Gemini CLI, each with a step-by-step setup. They connect through MCP. Connecting Claude and ChatGPT by signing in is coming with the hosted pilot.
Your agent's chat asks you to allow one action at a time, and the decision scrolls away with the chat. Agent Process asks the right person to approve the result, with the evidence attached, from an email link if they like. The run and the decision stay on record, and the same process works in any supported agent.
No. Describe the process to your agent or draw it on the canvas, connect a database or import an API definition, write your email templates and choose who signs off. During the pilot, we build your first process with you.
No. Passwords and keys stay encrypted with Agent Process and never reach the agent. Databases are read-only, and an API action that changes data works only after an admin enables it. These limits cover work routed through Agent Process; tools your agent has elsewhere are outside our control.
Jev is the decision AI from TypeSafe, built into Agent Process. It makes judgement calls inside a process: sorting a request into your categories, scoring work on your scale, or checking whether a statement holds. When it's unsure, a person decides, and agents can't change its answers.
The Model Context Protocol (MCP) is the open standard AI agents use to connect to tools and data. Agent Process gives your agents one MCP server with your processes, data, email templates and approvals.
It's free during the pilot. Join the waitlist and we'll invite you in small groups and help you build your first process. We agree any pricing with you before a paid plan starts.
Bring one process.
Tell us the job. When you’re invited to the free pilot, we set it up with you on a call and run it with your agent.