Agents and automations · Ultimate

Describe the flow. Approve the work.

Say what should happen in plain words and the copilot wires the blocks. Agents chase invoices, write outreach, screen candidates, read receipts and fix bugs, each inside limits you set, each stopping at a human gate when it matters.

Start 14-day trial See a flow being builtIncluded in Ultimate · €58 per seat per month

“Every Monday, list invoices more than 14 days overdue, draft a reminder for each, and let Sanne approve before anything goes out.”Understanding ✓Choosing blocks ✓Wiring▍

Trigger · ScheduleEvery Monday 08:00Europe/Amsterdam

Action · KairosList overdue invoicesdays_overdue > 14

Logic · LoopFor each invoicemax 50 per run

AI · Run AI agentInvoice chaserDraft reminder, tone: firm, Dutch

Human · ApprovalSanne approvesExpires after 48h → End

3 changes +1 ~2 −0 · copilot added the Approval gate and tightened the loopAccept allReviewDismiss

Copilot · Explain this flow · Test with sample data · Find problems · Optimize stepsRuns · Requests · Agents · Connections · Usage

Run inspector · last run, Mon 08:00Overdue invoices found6Reminders drafted6StatusWaiting for approvalSix reminders are waiting for Sanne. Nothing has been sent.

Block names, copilot commands and status sentences as they appear in the product.

Twenty blocks. Six families. One canvas.

The copilot picks them for you from a description; you can also drag them from the palette and wire them by hand. Every block carries its own test data and a schema check.

Triggers

EventScheduleWebhookManualPoll

Task, goal, CRM, finance, recruitment and messaging events; Slack, WhatsApp and Telegram messages; browser workflows.

GenerateRun AI agent

A single model call with a schema, or a full agent with its persona, knowledge, limits and autonomy settings.

Actions · Kairos

Action

Any of the ~350 workspace tools: create a task, post a journal, send an estimate, start a sourcing run, update a candidate.

Logic

ConditionSwitchLoopTransformWaitWait for event

Branch, iterate with a cap, reshape data, pause for a time or for something to happen.

Human

ApprovalGuardrail

A person decides before anything leaves the workspace; a guardrail stops a run that crosses a limit and logs why.

Advanced

Call workflowCodeNoteEnd

Compose flows from flows, drop into code when a block is not enough, annotate, and end explicitly.

Runs and requests

Every run reads like a sentence.

The inspector tells you what happened in words, not a log: “A pull request is waiting for review.” “The reminder was sent.” “Stopped at the guardrail: daily spend reached.” When an agent needs a person, the request lands in one inbox: approve, choose, type an answer, or talk to the agent.

Runs · today14 runs · €3.12

Chase overdue invoices
Six reminders are waiting for Sanne.Waiting for approval

Autofix · NL-INV-213
A pull request is waiting for review.Succeeded

Weekly digest
Reading 31 updates▍Running

Receipt reader
Stopped at the guardrail: daily spend reached.Stopped

Requests · 1 of 3ChoiceInvoice chaser asksHaven Logistiek has two open invoices. Send one combined reminder or two separate ones?One combined reminder●Two separate remindersAnswerTalk to the agentNothing has been sent.

PersonaKnowledgeGoalsLimitsScheduleAutonomyEscalationDelegation

Invoice chaserDraft onlyApprove firstAutoRead●Draft●Send external●Modify records●Spend●

Limits · €20 per day · 40 actions · 10 sends per channelRamp-up · 7 days at Draft only

When unsureAsk via interaction requestHand over to Finance lead agentStop and log

Start fromBlank agentDigest editorInvoice chaserOutreach writerReceipt reader

Agent designer

An agent is a job description with limits.

Persona, knowledge, goals, limits, schedule, autonomy, escalation and delegation, each on its own tab. The autonomy matrix says what the agent may read, draft, send, change or spend on its own. New agents ramp up for a week at draft-only before you widen anything.

Agents can hand work to each other and to the coding agent, and every model call lands in the audit with its cost.

Engineering · coding agent and code review

An error becomes a reviewed pull request.

Connect Sentry and GitHub to a goal. A model writes the spec, a coding agent fixes it in a sandbox or in your own GitHub Actions, a reviewer agent reads the diff, and a person merges. Choose how far it goes on its own.

Approve: the spec and the fix are prepared, a draft pull request is opened, and a person reviews before merge. The default.

SentryNL-INV-213VAT rounding differs on credit notes · 41 events

SpecRoot cause hypothesisPer-line rounding; fix proposal, forbidden paths, required checks

Coding agentClaude Code · Kairos sandboxFailing test → fix in vat.ts → 212 passed

Review laneReviewer agent2 suggestions, 1 question, no blockers

GitHubDraft PR · Fix credit note roundingA pull request is waiting for review.

AgentsClaude Code or CodexEight model providers; a budget per job; attempts capped.

Where it runsKairos sandbox or GitHub ActionsYour runners, your secrets; nothing leaves the repo you point at.

GuardrailsForbidden paths, required checksSpec panel on every task; governance audit of each run.

Connections

Wired to where the work already is.

Every connection is a provider with its own scopes and a test button. Flows start from a GitHub push, a Sentry alert, a Slack or WhatsApp message, a calendar event or an inbound email, and act through the same connections back out.

Kairos MCP · ~350 toolsKairos HTTPCoding agentMoira connectorGitHubGitHub ActionsSentrySlackGoogleMicrosoft 365SalesforceHubSpotEmailDriveCalendarWhatsAppTelegramLinkedInInstagramWebhooks, in and out

Governed like a team member, not a script.

One AI pool per teamTokens, actions and spend per day, with the assistant, agents and the coding agent drawing from the same budget. Usage is visible per flow and per run.

Approvals in one inboxRequests from any agent, the CRM's approval modes, recruitment screening replies and finance posting proposals, all in one place, with undo where an action already ran.

An audit of every model callWho asked, which model, what it cost, what it changed. Feature toggles for coding agents, external messaging and publishing, per workspace.

Automations, agents and the coding agent are part of Ultimate, €58 per seat per month. Quick research and the assistant's read-only skills are on every plan.

Describe the flow. Approve the work.

Every run reads like a sentence.

An agent is a job description with limits.

Agents can hand work to each other and to the coding agent, and every model call lands in the audit with its cost.

Governed like a team member, not a script.

One AI pool per teamTokens, actions and spend per day, with the assistant, agents and the coding agent drawing from the same budget. Usage is visible per flow and per run.

Approvals in one inboxRequests from any agent, the CRM's approval modes, recruitment screening replies and finance posting proposals, all in one place, with undo where an action already ran.

An audit of every model callWho asked, which model, what it cost, what it changed. Feature toggles for coding agents, external messaging and publishing, per workspace.

Automations, agents and the coding agent are part of Ultimate, €58 per seat per month. Quick research and the assistant's read-only skills are on every plan.

Describe the flow. Approve the work.

Twenty blocks. Six families. One canvas.

Every run reads like a sentence.

An agent is a job description with limits.

An error becomes a reviewed pull request.

Wired to where the work already is.

Governed like a team member, not a script.

Build your first flow from a sentence.

Describe the flow. Approve the work.

Twenty blocks. Six families. One canvas.

Every run reads like a sentence.

An agent is a job description with limits.

An error becomes a reviewed pull request.

Wired to where the work already is.

Governed like a team member, not a script.

Build your first flow from a sentence.