AI agents
An Unyo agent is a role, a hard-enforced toolbox and shared memory. Meet the 10 agents, see exactly what each one can and cannot touch, and learn why separate agents beat one general chatbot.
Bu sayfa senin dilinizde henüz mevcut değil - İngilizce sürümü gösteriyor.
An Unyo agent is three things at once: a role, a toolbox that is enforced in code, and a memory shared with every other agent.
That is the whole idea. Most AI products give you one assistant and change its instructions depending on what you ask. Unyo gives you ten coworkers, and what each one can reach is decided by the runtime, not by a paragraph of instructions the model may or may not follow.
The clearest example: Maya is the only agent with your inbox. Ashley cannot read your email. Not "Ashley has been asked not to" - Ashley's session never loads the email tools at all. There is nothing for her to call.
The three parts
1. A role
Each agent has its own name, colour, avatar, personality and domain. You open a conversation with the agent whose job matches the work - the same way you'd walk over to the right desk.
2. A toolbox, enforced twice
Every conversation builds a toolbox at request time. An agent only ever sees the tools its role is allowed to have, and the boundary is applied at two independent layers:
- The prompt layer - the agent is told what it owns and what it doesn't.
- The runtime layer - a per-agent denylist plus conditional mounting mean the forbidden tools are never loaded into the session in the first place.
So even a model that decided to misbehave has nothing to misbehave with. This is why we describe Unyo's limits as features: an enforced boundary is worth more than a well-behaved prompt.
3. Shared memory - the Neural Core
All ten agents read the same Neural Core. Not "we sync it" - the memory context is built without knowing which agent asked for it, so every agent gets a byte-identical view of your brand, products, contacts and knowledge.
Tell Maya "we never use emojis" and Ashley knows on her next run. There is also a shared workday digest: anything that happened across all your conversations in the last 18 hours is visible to whichever agent you talk to next.
Memory costs nothing
All 23 Neural Core tools cost 0 credits. Reading, searching and updating your own business memory is free, always. See Credits and billing.
The 10 agents
| Agent | Role | What they're for |
|---|---|---|
| Ashley | Social media | Social post drafts, your own post analytics, best-time and hashtag research |
| Maya | Email & calendar | Your Gmail or Outlook inbox - no other agent has it |
| Alex | Sales | Product cards, Shopify orders and abandoned carts |
| Riley | Product | Documents, spreadsheets, research, Slack/Notion |
| Tyron | Images & audiovisual | Image generation, editing, HD upscale, background removal |
| Sam | Customer support | Shopify order lookup for support answers |
| Lucy | Content writing | Research-then-write documents with real [n] citations |
| Blake | Data analysis | A real Python sandbox, plus Stripe revenue (read-only) |
| Ema | Personal development | The cross-agent "what's on my plate" snapshot |
| Steve | Websites | Building, previewing and publishing static websites |
Some capabilities are shared on purpose. Nine of the ten agents write real documents (Steve is the exception - he works in code, not prose), and all ten can build you a multi-tab Excel workbook as a downloadable card.
Maya's boundary, in plain terms
The email tools only load in Maya's conversation and in scheduled automations. Every other agent's session is built without them. If you ask Ashley to check your inbox, she will hand the task to Maya - because she genuinely cannot do it herself.
How this differs from a chatbot
| A general chatbot | An Unyo agent |
|---|---|
| One assistant, instructions swapped per topic | Ten roles, each with a toolbox built at request time |
| "Please don't touch X" is a sentence in a prompt | X is never loaded into the session |
| Writes you text about the work | Produces the artefact - an editable draft card, a document, a spreadsheet, a website |
| Forgets between chats | Every agent reads the same Neural Core |
| Answers from training data | Searches and reads the live web, and grounds answers in your own Neural Core |
The last row matters more than it looks. When Ashley gives you a post, you get an editable draft card - caption, hashtags, and a generated image - not a wall of text you have to copy out. When Blake does maths, he runs actual Python and hands you the chart and the file. The artefact is the product.
How to use them
- Open a conversation with the right agent. Pick from the sidebar. If you're not sure, start anywhere - the agent will hand off.
- Ask in plain language, in any language. You don't need to name tools.
- Watch it work. Most agents declare a plan, then announce each step as they start it, live in the transcript. Expand a finished step to read what it actually produced - the real caption, the real subject line.
- Edit the result. Drafts, documents and spreadsheets are editable cards, not final answers.
- Confirm anything irreversible. Sending an email, posting to Slack, deleting a calendar event
- these surface a confirmation card. Nothing leaves your account until you click.
Ashley, draft an Instagram post announcing our new autumn collection. Use our brand tone from the Neural Core.
When one agent needs another
Ask an agent for something outside its role and it hands the task to the right coworker automatically.
Delegation is real, not a figure of speech. The coworker opens its own conversation, runs the full pipeline there, and you get a notification the moment the hand-off happens. Open that conversation to see the work.
Two things worth knowing:
- The hand-off is a hand-off, not a merge. The agent that delegated does not receive the coworker's output and will not pretend it did the work itself. You read the result in the coworker's conversation.
- Steve is the exception. No agent can hand a website to Steve. If you want a site built, open a Steve conversation yourself. Steve can delegate out - he just can't be delegated to.
There's also a loop guard: a conversation that was itself created by a delegation cannot delegate again. One hop, no runaway chains.
Two ways to hand off work
Delegation is immediate - the coworker starts now. There's also a slower route: an agent can leave a task in the Neural Core addressed to a coworker, who picks it up in their next session. Use delegation for "do this now", tasks for "get to this".
Languages
Set your agent language once and every agent replies in it, whatever language you type in. 11 agent reply languages are available (English, French, Spanish, German, Italian, Portuguese, Turkish, Arabic, Chinese, Japanese, Russian). You can also just ask an agent mid-conversation to reply in another language for a single answer.
Your agent language is separate from your interface language - see Languages.
What agents will not do
These aren't warnings, they're the design.
- They don't send or publish on their own. Email sends, Slack posts and calendar deletions are confirmation-gated. Website publishing is explicit and priced - Steve never publishes on his own initiative.
- Recurring automations draft, they don't send. An automation runs as the agent you set it up with and prepares something for your review. Unattended sending does not exist in Unyo. See Automations.
- They don't touch what they can't reach. Stripe is Blake's alone and is read-only - there is no charge, refund or payout tool anywhere in the product. Shopify is read-only too.
- They don't claim credit for a coworker's work. See the delegation section above.
Are these really ten different agents, or one chatbot with ten names?
Ten. The difference is enforced in code, not in the prompt: each conversation builds its toolbox at request time from the agent's role, and forbidden tools are never mounted into the session. Ashley's session has no email tools to call. Blake's is the only one with Stripe. Steve's is the only one with the website tools. It's an access boundary, not a personality setting.
Which agent should I start with?
Whichever matches the job - Maya for email, Ashley for social, Lucy for writing, Blake for numbers, Steve for a website. If you pick wrong, the agent will hand the task to the right coworker and notify you. The one exception is Steve: websites are only built in a Steve conversation, so open that one yourself.
Do the agents share what I tell them?
Yes - that's the point. All ten read the same Neural Core, and it's architectural rather than a convention: the memory context is built without reference to which agent requested it, so every agent sees the same thing. Tell one agent a preference and the others have it too. What they do not share is access: shared memory, separate toolboxes.
Can an agent send an email or publish a post without me?
No. Sending is confirmation-gated, and unattended send does not exist in the product - not as a setting, not per workflow. Recurring automations prepare a draft and stop. If something is going out under your name, you clicked.
Can two agents work on the same answer together?
Not in one reply. Delegation opens a separate conversation for the coworker, who does the work there and notifies you. You'll read the result in that conversation. The delegating agent tells you it handed the task over rather than pretending the output is its own.
Troubleshooting
An agent says it can't do something you're sure Unyo does. It's probably not that agent's job. Ask it to hand the task over, or open a conversation with the right coworker directly. The capability lists in the table above are the real boundaries.
An agent asks you to connect an account. Agents check which platforms you've connected before acting. If a capability needs an account you haven't linked, you'll get a Connect card instead of a failure. Click it to link the account, then ask again.
A step failed and you were charged. You weren't. Every paid tool refunds automatically when it errors, and an hourly janitor refunds anything stranded. See Credits and billing.
A delegated task looks like it vanished. It's in the coworker's own conversation, not the one you were in. Check your notifications - one fires at the moment of hand-off.
Next: How agents work · Delegation · The Neural Core