> ## Documentation Index > Fetch the complete documentation index at: https://learn.actionist.ai/llms.txt > Use this file to discover all available pages before exploring further. # Agents > Create, configure, and supervise AI employees — the "who" behind every workflow Actionist runs.
YOUR AI WORKFORCE

Hire your first AI employee in under an hour.

Give each agent its own role, tools, memory, and channels. Agents keep the business running while you focus on the interesting problems.

Maya · Marketing Receptionist Customer Support Sales Assistant Executive Assistant
The idea

Not a chatbot. An employee.

A chatbot waits to be asked. An Actionist agent has a role, a routine, and the tools to finish the job.

It responds when you type, forgets when you close the tab, and can't touch your tools, files, or customers. Every task starts from zero. It owns a role, remembers across conversations, uses your real tools through MCP, runs on a schedule, and escalates only when it's genuinely stuck — not when it's unsure. Agents are AI employees. They have working hours and daily routines, proactively pick up work from their tools, collaborate with each other through delegation, remember context across conversations, and escalate when stuck, not when unsure.
Follow one example

Meet Maya.

To make this concrete, we'll build one agent and follow her through the whole page — from a blank studio to a full working day.

Owns marketing operations end to end: checks analytics every morning, monitors competitors, drafts content for review, and reports daily. Never publishes without sign-off, never approves ad spend over a threshold without you. By the end of this page she's a real, scheduled, supervised employee.
Anatomy

What makes an agent.

Every agent is three things: who it is, what it can do, and how it's governed. Assemble Maya layer by layer — click through the three layers and watch her agent card build up.

MayaMarketing Manager
identity · instructions · soul · model
Owns marketing ops end to end Never publishes without sign-off Voice: data-informed, zero hype Model: preset configured
tools · skills · memory · channels
FilesComputer useAnalytics MCP Skill: brand-contentMemory: CRM readTelegram bot
approval · collaborators · schedule
Guarded auto Delegates to: Designer agent Weekdays 09:00 Trigger: webhook
draft · has an identity, can't act yet capable · armed with tools, ungoverned deployable · scoped, supervised, scheduled
Name, role, and a one-line description — what teammates see in Slack, Telegram, and the agent directory. Maya, Marketing Manager. The system prompt the agent runs under: what it always does, what it never does, and the tone it uses. Short, vivid prose the model uses as a voice reference. This shapes personality, not behavior. Pick the brain: choose the model preset that fits the task. Heavier work runs on more capable presets.
Built-in actions — files, terminal, computer use, messaging, sub-agents — scoped on or off per agent. Portable behavior packs: instructions plus a curated tool subset. Install from the marketplace, toggle per agent. Permissioned buckets — company-wide context, agent-private knowledge — with read, write, or no access each. Where the agent talks: the desktop, a Telegram bot, a Slack workspace. Each agent can own its own bot.
How much the agent does on its own versus pausing for your sign-off. Supervision you dial in — the full dial is below. The other agents this one may delegate to, or hand a whole conversation off to. One-off and recurring jobs plus event triggers. The agent shows up to work without being asked.
Agent Studio · Maya · Marketing Manager
✎
Maya
Marketing Manager
Identity
Tools
Skills
Memory
Channels
Collaboration
Schedule
Triggers
Runs
Settings
Name
Maya
Role
Marketing Manager
Instructions
You are Maya, the Marketing Manager. You own marketing operations end to end…
Soul
Creative, data-informed, genuinely excited about growth…
Live preview · Readiness
✓ Name · under 30 chars
✓ Role · specific
✓ Instructions · 412 chars
✓ Soul · voice present
Deployable
Hi! I pulled yesterday's analytics — signup rate is down 24%. I've flagged variant B as the likely cause. Ready to draft a fix for your review.
The craft

Write the brain.

Instructions decide what an agent does. Soul decides how it sounds. Keep them separate and both stay sharp. Here's Maya's:

```md Instructions.md theme={null} # Role You are Maya, the Marketing Manager. You own marketing operations end to end. ## You always - Check analytics every morning and flag anomalies with context. - Draft content against the brand guidelines before anything ships. - Report daily: lead with the metric, then the recommendation. ## You never - Publish without explicit human sign-off. - Approve ad spend over $500 without approval. - Report a vanity metric without the context that makes it mean something. ## Escalate when - A metric moves more than 20% day over day. - A campaign needs budget you're not cleared to spend. ``` ```md Soul.md theme={null} ## Voice Creative, data-informed, and genuinely excited about growth — never hype for its own sake. ## Style Lead with the number. Bullet the status. Always end with a recommendation, not just an observation. ## Boundaries Energetic, not overwhelming. Confident, not robotic. You bring the idea and the evidence in the same breath. ``` If the description fits in a tweet, the agent behaves consistently. If it takes a paragraph, split it into two agents. A narrow remit beats a clever one.

The readiness checklist, live.

The studio's Identity tab grades your draft against four checks before it calls the agent deployable. Toggle between a first draft and Maya's final config to see what each check wants:

Name ✓ Maya — under 30 characters ✓ Maya — under 30 characters
Role ✗ assistant — too generic; the check wants a specific role ✓ Marketing Manager — specific, not "assistant"
Instructions ✗ "Help with marketing." — 21 characters; needs over 200 ✓ the full always / never / escalate block above — 412 characters
Soul ✗ empty — the check wants a voice of over 40 characters ✓ "Creative, data-informed, genuinely excited about growth…"
1 of 4 checks passing — the studio flags this agent as not ready. 4 of 4 checks passing — deployable. This is the bar every agent should clear before its first scheduled run.
The studio

The studio, tab by tab.

Whichever path you took, every agent is tuned in the same place. Click a tab to see what it controls.

Agent Studio · Maya
Identity
Name, role, description, Instructions, and an optional Soul. A live preview and readiness checklist confirm the agent is deployable — it checks: Name under 30 characters; Role is specific (not "assistant"); Instructions longer than 200 characters; Soul describes a voice (over 40 characters). You can also personalize the avatar: click the pencil badge on hover to choose a shape and color applied workspace-wide.
Maya · Marketing Manager
Tools
Built-in tools — file management, terminal, computer use, messaging, memory search, sub-agent delegation — always available but toggleable per agent. MCP servers and skills add more. The global approval mode sets the default; per-tool overrides let you trust the safe tools and gate the risky ones.
approval mode lives here
Skills
Portable behavior packs — instructions plus a curated tool subset. A behavior you perfected on one agent becomes reusable across your whole workforce. Enable from the Skills page globally, then wire here per agent.
key icon appears when credentials are needed
Memory
Permissioned per bucket — global, company info, products, CRM, policies, and any custom buckets — each set to Read, Read and write, or None. Memory instructions tell the Memory Agent what's worth keeping, so the agent sharpens over time instead of forgetting.
brain icon
Channels
Bind the agent to a Telegram bot or a Slack workspace. Each agent can have its own bot, or share a company-level Slack app and route by mention name. Tokens are stored encrypted in your workspace vault — only this agent can read them.
encrypted vault tokens
Collaboration
List the agents this one can delegate to. A manager agent can hand a task to a specialist, or hand off the whole conversation so the specialist talks to the user directly. Recent delegations and hand-offs show here so you can trace who did what.
delegate\_task and hand\_off tools
Schedule
Give the agent a heartbeat. Scheduled jobs run recurring work — Maya's 9am analytics check — using familiar cron expressions or plain-language recurrence. Set the timezone once; Actionist handles daylight saving.
0 9 \* \* 1-5 · weekdays at 09:00
Triggers
Fire the agent on external events — an inbound webhook, a new Slack message, an email arriving, or another agent completing a task. Triggers let the agent react to the world without polling.
webhook · email · agent-complete
Runs
A per-agent run history. Every run this agent has taken — status, start time, duration, and tool calls — scoped so you can audit its track record without wading through workspace-wide logs. Use it to spot patterns: why a scheduled job failed last Tuesday, which tool call took the longest.
status · duration · tool calls
Settings
Enable or disable the agent, set a Critical Reminder (up to 400 characters) prepended to every turn, and find the danger zone — deleting an agent requires typing its name and can't be undone. If running locally with a paired cloud runtime available, a Move to VPS button migrates memory, skills, and schedule to the cloud in one step.
critical reminder · danger zone · move to VPS
Get started

Two ways to create one.

Open Agents in the sidebar, hit New agent, and pick the path that fits.

Describe what you need; the assistant configures tools, memory, and channels as you chat. Best when you know the job, not the settings. Watch Maya get built:

Agent creator
youA marketing manager that checks analytics every morning, drafts content for review, and reports daily. Never publishes without my sign-off.
designing identity
aiMeet Maya, Marketing Manager. I've drafted her instructions with your sign-off rule as a hard "never". Tools she'll need:
Analytics MCPFilesComputer useTelegram
youAdd a \$500 ad-spend approval threshold.
aiAdded under "You never": approve ad spend over \$500 without approval. Anything else?
saved · maya appears under My Agents
aiDone — open her in the studio anytime to fine-tune.
Start with a blank agent in the studio and configure identity, tools, and skills yourself. Best when you know exactly what you want. Every tab from the explorer above is yours from the first second — and the readiness checklist keeps score as you go.
Already know the role? Open the **Templates** tab from the Agents list view to browse 8 pre-wired role templates and open one directly in the studio — no chooser needed.
{/* TODO(image): create-chooser — IMAGE SPEC: The 'Create agent' chooser modal showing the two options side by side: 'AI-guided setup' (with the Recommended badge) and 'Manual setup', each with its icon and one-line description. — replace this CSS mock with the real screenshot */}
Create agent
Recommended
◆
AI-guided setup
Describe the job — the assistant configures tools, memory, and channels as you chat.
⊞
Manual setup
Start with a blank agent in the studio and configure each setting yourself.

The recommended path: build Maya with AI-guided setup

Choose AI-guided setup. The Agent creator opens — "Chat to design your agent — refine identity, tools, and skills step by step." "A marketing manager that checks our analytics every morning, monitors competitors, drafts content for review, and reports to me daily — and never publishes without my sign-off." The assistant proposes Maya's name, role, instructions, tools, and memory access. Tap a suggestion or type your own until it's right. Maya is saved and appears under My Agents. Open her in the studio anytime to fine-tune tools, skills, schedule, and memory.
Trust

Supervision is a dial, not a switch.

Start every agent cautious. Loosen the leash one notch at a time as it earns your trust. Select a mode below to see which action classes auto-execute and which pause for your sign-off.

Suggest — Maya proposes every tool call and you run them manually. Maximum oversight. Start here while you learn what an agent will do.
Ask each time — Maya requires your approval before every tool execution. Nothing happens without an explicit yes.
Guarded auto — Maya runs safe reads and writes automatically; sensitive writes and anything destructive pause for your sign-off. The sweet spot for most production agents.
Full auto — Maya runs every tool without prompting. Fast and hands-off — for agents you've already learned to trust. Destructive actions still always prompt; this floor cannot be removed.
Action class behavior in this mode
Read manual prompt auto-execute auto-execute Search memory, read files, fetch URLs
Safe write manual prompt auto-execute auto-execute Draft saves, memory writes, internal notes
Sensitive write manual prompt prompt auto-execute Send email, post to Slack, create CRM records
Destructive manual prompt always prompt always prompt Delete files, cancel subscriptions, drop records
Even in Full auto, any action flagged as destructive always prompts for approval. This floor cannot be removed — so you can never accidentally authorize a silent delete.
New agents default to Full auto. If the agent can write to external systems — send email, create records, modify files — change the approval mode to Guarded auto immediately after creation. Guarded auto approves safe reads automatically and surfaces writes and destructive operations for your sign-off.
Templates

Start from a role, not a blank page.

Each starter template comes prewired with identity, tools, and skills. Adjust everything before saving.

Popular

Greets visitors, answers FAQs, and routes people to the right specialist.

 Popular

Handles tickets end to end and escalates the ones that need a human.

 Qualifies leads and runs the follow-ups that usually slip through the cracks. Keeps accounts, contacts, and deals clean and up to date. Provides engineering oversight and keeps delivery on track. Generates and debugs code against your stack and conventions. Automates the recurring process work that runs the business. Triages email, manages the calendar, and coordinates across people.
{/* TODO(image): template-gallery — IMAGE SPEC: The starter-template gallery: a grid of template cards (Receptionist, Customer Support, Sales Assistant, CRM Manager, Technical Lead, Developer Agent, Operations Assistant, Executive Assistant), each with its icon, name, and one-line summary, with the STARTER and POPULAR tags visible on a couple. — replace this CSS mock with the real screenshot */}
Templates 8 role templates
STARTERPOPULAR
Receptionist
Greets visitors, answers FAQs, routes to the right specialist.
STARTERPOPULAR
Customer Support
Handles tickets end to end, escalates when a human is needed.
STARTER
Sales Assistant
Qualifies leads and runs follow-ups that slip through the cracks.
STARTER
CRM Manager
Keeps accounts, contacts, and deals clean and up to date.
STARTER
Technical Lead
Engineering oversight and keeps delivery on track.
STARTER
Developer Agent
Generates and debugs code against your stack.
STARTER
Operations Assistant
Automates recurring process work that runs the business.
STARTER
Executive Assistant
Triages email, manages calendar, coordinates across people.
More roles

More seats you can fill today.

Maya is one example. The same pattern — role, tools, schedule, supervision — works for every function in the business.

Dario · Sales
Each morning Dario pulls the CRM, identifies deals with no activity in 5+ days, drafts personalised follow-up emails for each one, and queues them for your review. Replies that come in during the day are logged against the right deal automatically — no manual data entry.
saves \~3 hrs/day on pipeline hygiene
Priya · Support lead
Priya watches the support queue, classifies and triages each ticket against your policy docs, resolves the clear-cut ones directly, and surfaces the edge cases with a recommended action. Agents on Guarded auto: she never changes a subscription without your sign-off.
saves \~5 hrs/day on first-touch resolution
Felix · Finance and Ops
Felix runs a weekday close: pulls invoices from the inbox, reconciles against the ledger, flags any mismatch above a set threshold, and drops a one-page cash-flow summary into the shared drive before 08:00. Nothing is posted without an approval gate.
saves \~2 hrs/day on reconciliation and reporting

One agent rarely owns the whole outcome.

A manager agent delegates to specialists, runs them in parallel, and assembles the result. The pulse below shows delegation flowing from you through the manager to three specialist agents.

delegation flow
YouSend the task. The manager handles the rest.
Manager agentBreaks the work into parallel sub-tasks.
delegate\_taskEach specialist receives scoped instructions and tools.
Assembled resultThe manager merges the outputs and reports back.
Developer agent Generates and tests code against your stack.
QA agent Runs acceptance tests and flags regressions.
Research sub-agent spawn\_subagent — spawns with a fork context for long-running analysis.

A product-manager agent can take a feature from spec to shipped: it delegates the build to a developer agent, the testing to a QA agent, and the deploy to an operations agent — coordinating the whole team in parallel. A solo founder ships like a full team.

Before and after

Tuesday, before and after Maya.

The same marketing responsibilities — two very different mornings.

Before · the manual grind
08:55Open four browser tabs, wait for dashboards to load.
09:10Copy yesterday's numbers into the weekly spreadsheet by hand.
09:28Write standup update from memory — three Slack threads deep.
10:05Draft a launch post, then spend 20 minutes checking brand-voice guidelines.
11:30Realize the anomaly from Monday was never followed up on.
14:00Pull the campaign report — discover it needs a manual CSV merge.
17:00Forget to escalate the ad-spend decision before end of day.
After · what you actually do now
Review Maya's 09:00 analytics brief and flag one decision that needs you.
Approve or tweak the launch post draft she held for sign-off.
Respond to the escalated ad-spend question she surfaced by 17:00.
The payoff

Maya's Tuesday.

Scheduled, supervised, and on-brand. Nobody tells her to start.

schedule · maya · tuesday
schedule · daily-analytics · 09:00
09:00Pulled yesterday's analytics. Signup rate down 24% — flagged with likely cause (landing page variant B).
09:01Posted to #marketing-daily. Awaiting your response before any campaign changes.
schedule · standup · 09:30
09:30Standup sent: what shipped, what's blocked, what she's picking up today — in her own voice, by design.
schedule · midday-sweep · 13:00
13:00Checked campaign performance. Drafted launch post for Product Hunt — held for your review, never auto-published.
schedule · eod-report · 17:00
17:00Day summary sent. Budget decision escalated — ad spend approval needed for Tuesday push.
17:00Everything else wrapped. No surprises.
Nobody tells her to start. The schedule fires automatically — she shows up, does the work, and reports back.
Never auto-publishes. Approval mode is set to Guarded auto: content drafts wait for your explicit sign-off before anything ships.
Escalates budget calls. Instructions say never approve ad spend over \$500 without you — so she doesn't, and flags it clearly.

That's one agent. Now imagine the support, sales, and operations seats filled too.

The same agent runs everywhere.

Pick an agent from the chat selector and work side by side. Every step, tool call, and screenshot streams live. Bind a bot and the agent answers your team where they already are — DMs, channels, threads, mentions. Recurring jobs and event triggers mean the agent shows up without anyone asking it to. Headless Runtime keeps agents firing when your laptop is closed — local, cloud, or hybrid.
Quickstart

Your first agent, in five minutes.

Download for Mac or Windows from [the dashboard](https://app.actionist.ai). Sign in with email or SSO. In the sidebar, open Agents. You'll see Recently used, My Agents, and Templates. Hit New agent and choose AI-guided setup — the fastest way to a working agent. "Summarize my unread Slack DMs every morning and post the digest to my #daily channel." Answer the questions and save. Chat with the agent from the desktop, or bind it to a Telegram or Slack bot and let it run.
Ready when you are

Build your first agent today.

Describe the job. Actionist wires up the tools, memory, and channels — you supervise.

Your agents do exactly what you allow — no more.
Scoped permissions · Approval gates · Full audit trail.

Keep going

Next steps.

Turn a behavior into a reusable, publishable pack every agent can load. Give agents permissioned recall — company context that sharpens over time. Chain agents and steps into repeatable, schedulable processes. Put an agent on Telegram or Slack so it works where your team already is.