Agents that follow
your process.
Write the playbook in plain language. The agent runs it step by step, calls only the tools you gave it, and proves it before it ever talks to a customer.
Chat agents
Refund handler
Handles return and refund requests end to end, inside the policy window.
Order status
Answers where-is-my-order questions and escalates anything late or damaged.
Internal agents
Invoice chaser
Chases overdue invoices every morning and flags the ones that need a human.
Ticket triage
Reads every new ticket, sets priority, and routes it to the right group.
Knowledge keeper
Answers product questions from the knowledge base when someone @mentions it.
Five hires. No onboarding.
Each one gets a job, a set of tools, and an access level, the same three things you give a person. Then it starts itself.
Refund handler
Customer message- Refund requests
- 4 toolsets · 11 tools
- Support profile · Edit on Tickets
Order status
Customer message- Where's my order
- 3 toolsets · 7 tools
- Support profile · Read on Orders
Invoice chaser
Schedule · 08:00 daily- Overdue invoices
- 3 toolsets · 9 tools
- Billing profile · Edit on Invoices
Ticket triage
Event · Ticket created- Triage & route
- 4 toolsets · 10 tools
- Support profile · Edit on Tickets
Knowledge keeper
@mention in a comment- No procedure
- 2 toolsets · 5 tools
- Read-only everywhere
Not every agent needs a playbook. One with no procedure just answers, using the persona and the tools you gave it. No branching, no extra model call.
The agent doesn't decide the steps.
You do.
A procedure is a document you write in plain language. Inline badges make it executable: tools, code, branches, and where it ends. Pick an agent to watch one run.
You handle returns and refunds for Alder Supply. Be warm and brief. Never promise money back before you have checked the delivery date.
Step-by-step playbooks the agent follows for specific situations.
The customer wants money back on something they already received.
The item arrived broken, or the wrong item shipped.
The customer lost the return label or it has expired.
Waiting for a message. One of these will be picked, not all of them.
Only the tools you gave it
Every agent carries an allow-list, down to individual tools. Mention a tool in the prompt and it switches itself on; nothing else is ever in reach.
It can't outrank you
Access is snapshotted onto the published version and clamped by whoever published it. Run one on a colleague’s behalf and their access becomes a ceiling: delegation narrows, never widens.
Seven ways to start
A schedule, a record change, an installed app, an inbound webhook, an @mention, a ticket assignment, or a direct message.
Ship agent changes like code.
A prompt tweak is a deploy. Run the suite, read the diff, and know exactly what you fixed and what you broke, before a customer finds out.
Not a vibe check.
Simulations run the real agent loop against a synthetic customer, with every tool stubbed at the edge.
The real loop, mocked at the edges
Selection, stepper, and engine are the ones production runs. Only tools are stubbed, and an un-stubbed call errors the run instead of touching anything. No CRM writes, ever.
Assertions that don't argue
This tool with these arguments. This terminal outcome. This field ended up here. This variable equals that. Judged criteria only for prose.
error is not failed
A judge that breaks, a mock that's missing, a run that times out: those are errors, not verdicts. Nothing passes quietly.
Pinned or draft
Pin the published version and it's a regression gate. Point it at the draft and it's the iteration loop. A tool whose schema moved under a pinned run fails loudly.
Agents don't run alone.
They read the same records, answer from the same knowledge, and start off the same events as everything else in the workspace.
- [01]Kopilot
The same engine, driven by you instead of a trigger. It also builds the agents.
- [02]Knowledge base
What an agent answers from, scoped per agent and searched at run time.
- [03]Workflow
For the deterministic parts. An AI node hands a turn to an agent mid-graph.
- [04]Sequences
Timed reminders that need no judgment. Cheaper than an agent, on purpose.
- [05]Ticketing
Where triage writes, where assignment fires, and where a handoff lands.