How It Works
It loops, instead of answering
A chatbot produces one response and stops. If that response is wrong, incomplete, or based on a page that didn't load, that's your problem to notice.
AI Partner works differently. It keeps going round a cycle until the thing you asked for exists:
You describe what you want
│
▼
Break it into steps
(and work out which depend on which)
│
▼
┌──────────────────────────────────────┐
│ 1. Decide what to do next │
│ 2. Do it │
│ 3. Look at what came back │
│ 4. Adjust the plan │
│ └── repeat ─────────────────┐ │
└────────────────────────────────┬─┘ │
│ └─────┘
▼
Check it against what you asked for
(Does the file exist? Was the message sent? Is the data right?)
│
▼
Hand it over
The important part is the second-to-last box. "Done" means checked, not "the agent stopped talking." If you asked for a deck and it produced a summary, that's an incomplete run — and it's treated as one.
Three ways it runs
Chat
Quick questions, short tasks. It answers in one turn, streaming back as it goes. Chat is deliberately limited to things that only read — it won't write files or act on your accounts here, so you can ask freely.
You: What's the current price of RELIANCE.NS? AI Partner: ₹2,847.50 as of 11:32 AM IST
Goal
Multi-step work. The full loop: planning, doing, checking, correcting, and running independent parts in parallel. Start one with the rocket icon, or just describe something that obviously takes several steps.
You: Research the top 5 cloud providers by market share, write a comparison table, and build me a deck for Monday's board meeting.
Typically 30–120 seconds, ending in a file.
Meeting
A meeting link is its own thing. AI Partner opens the meeting, joins, listens and responds — see Meeting Proxy.
You: Join https://teams.live.com/meet/… as "Alex", focus on our Q2 roadmap discussion.
You rarely have to choose. A request that plainly needs several steps is treated as a goal; a question is answered as chat. The rocket icon is there for when you want to be certain.
How it plans
A goal is broken into steps with a kind and an order, because some can't start until others finish:
| Kind of step | Example |
|---|---|
| Find something out | "Find the top 5 cloud providers by 2025 market share" |
| Work something out | "Calculate year-on-year growth from these numbers" |
| Produce something | "Create a deck with the comparison table" |
| Send something | "Send the file to my Telegram" |
| Confirm something | "Check the file exists and has at least 5 slides" |
Steps that don't depend on each other run at the same time — several strands of research in parallel rather than one after another. See Agent Pool.
What it can reach
AI Partner has around 160 abilities grouped by what they're for. It sees what's available before each decision and picks — you never name the tool.
| Area | What it can do |
|---|---|
| Web | Search, read pages, and drive a real browser when a page needs clicking |
| Running code | Run scripts safely to process data, transform files and do calculations |
| Files | Read, write and organise files — in its own workspace, or on your computer |
| Producing documents | Decks, spreadsheets, documents, reports, PDFs, images and video |
| Communication | Email, Slack, Telegram and the other messaging channels; phone calls where you've allowed them |
| Your tools | GitHub, Jira, Notion, Confluence, HubSpot, Stripe, Drive, S3 and more — see Integrations |
| Your knowledge | Your documents, research papers, feeds |
| Its own memory | What it already knows about you, your history and your people |
| Delegation | Handing parts of a job to other agents and gathering the results |
Connect a service and its abilities simply appear. Nothing else to configure.
How it picks a model
Different work suits different models, and paying frontier prices to classify a message is a waste. AI Partner routes each piece of work to an appropriate model automatically:
| The work | The kind of model |
|---|---|
| Hard reasoning and planning | The strongest model you've connected |
| Writing code | A coding-specialised model |
| Quick classification | A fast, cheap one |
| Anything involving looking at a screen | One that can read images |
| Questions needing current sources | A search-grounded model with citations |
Override any of it in System → Model Routing, or pin one message to one model by prefixing it with @model-name.
How it recovers
Most of what separates a demo from something you'd rely on is what happens when a step fails.
When something it wrote fails, it reads the actual error, works out what was wrong, and tries a corrected version — a bounded number of times, not forever.
Repeating the same action with the same input and getting nowhere is detected as being stuck. It breaks the pattern and re-plans instead of looping.
A goal can't run indefinitely. When it reaches its limit it hands over what it has, says clearly what's missing, and doesn't pretend otherwise.
When it asks you
Some things shouldn't be decided by an agent. In those cases it pauses and asks:
| Situation | What you'll be asked |
|---|---|
| Permission | "Should I delete these 50 files?" |
| A credential | A sign-in or a one-time code it can't and shouldn't have |
| An ambiguous instruction | Your goal has two plausible readings, and they lead somewhere different |
| Cost | The work is about to cost more than your threshold |
| Something destructive | Writing to something that matters, irreversibly |
| A blocked page | A CAPTCHA or login wall — it shows you the live page so you can clear it |
| Its own uncertainty | The result isn't good enough to hand over as an answer |
You'll be asked in the web app as a card in the conversation, and on your messaging channel as a message with approve / edit / skip buttons — so a run isn't stalled just because you're away from your desk.
Answer, and it picks up exactly where it paused. It doesn't start over.
What it may do without asking at all, what it must ask about, and what it may never do are yours to set, in your authority policy.
A run, end to end
"Research NVIDIA's Q1 2026 earnings and make me a slide deck."
1. Recognised as a goal, not a question
2. Broken into steps:
find the earnings figures
find analyst reaction
work out the key metrics
build the deck
confirm the deck is real
3. Checked for a saved skill that already does this → none yet
4. Working:
searched for the earnings release
read the release and pulled out the numbers
searched for analyst commentary
processed it all into structured figures
built the deck
opened the file to confirm: 8 slides, data present
5. Checked against what you asked for → met
6. Saved what worked as a reusable skill
7. File in Generated Files, download link in the chat
8. Sent to your Telegram, if you've connected it
About 45–90 seconds.
Ask for a similar deck next month and step 3 finds the saved skill — so it runs the known solution instead of working it out again, and finishes in seconds. That's Skills.