Agents
Growing your agents
An agent is a member of staff on your AI team. Seventeen of them ship with the app — commander, writer, reviewer and the rest — and you can open any one up to see what it judges by, what it knows and how it speaks, then grow it as you work. Whatever you teach it shows up in its answers from then on.
Opening it
In the sidebar, go to More → Agents.
Pick a role from the list and its screen opens.
Roles are grouped into You, Basics, Creation, Operations, Admin and Yours. The number beside a name is how many entries that role currently holds.
＋ on the list heading creates a new role. Adding one and growing it happen on this same screen, so there is nothing separate to learn once you have made one. ⤓ beside it takes in a role someone has handed you (handing roles over).
An agent is a person (who handles it), a skill is a procedure (how it's done), and knowledge is what's known (the facts behind it). Your digital twin is you. It sits at the top of the same list, under You, and is grown on the same screen, but it is treated as separate from the role agents. Turn on Overlay your persona in the settings and your own way of deciding is layered on top of the role's.
Reading the screen
The screen is split in two.
| Side | What it's for |
|---|---|
| Left: Agent knowledge | What the role holds — where everything accumulates and gets tidied up |
| Right: Assistant | Where you talk to it. Requests, teaching and handing over files all happen here |
What you say on the right surfaces on the left as a proposal. Whether it's adopted is always your call.
The four things that grow
Agent knowledge on the left is divided into four. Each has a cap, because past a certain point the AI can no longer tell what it should prioritise.
| Section | What goes in | Cap |
|---|---|---|
| ⚖ Judgment rules | What to put first when it's a close call | 20 |
| 📒 Knowledge | Facts, rules and figures worth remembering | 30 |
| 💬 Style | Tone, the order things are said, length | 10 |
| 🧠 Approach | The shape of how the work gets done | 10 |
One entry says one thing. Rather than pasting in a long block, add it grain by grain. Every entry carries a why (the reasoning) and where it came from — Default, added directly, taught in chat, or from a file.
All 17 roles ship with entries already in place. The reviewer, for instance, comes with the 56 permitted efficacy claims under Japan's Pharmaceutical and Medical Device Act, plus the Act against Unjustifiable Premiums and Misleading Representations and Japan's rules on undisclosed advertising (stealth marketing). The writer comes with house style rules, sales with the points to confirm in a meeting, and legal with Japan's Freelance Act and its subcontracting fairness act. You are not starting from an empty page.
Four ways to grow one
With any of them, the AI never rewrites anything by itself. A proposal appears, and it only takes effect once you adopt it.
1. Add it directly
Press ＋ on a section heading and an input field opens on the spot. An entry that already exists can be rewritten by clicking it.
2. Teach it in conversation
Tell the assistant on the right “from now on, do X” or “that's wrong, it should be Y” and the proposal appears as a card. Adopt commits it, Skip discards it.
You can also just give it work. It answers in character, and when it applies one of its judgment rules it adds a short (rule: …) note in the body, so you can see with your own eyes that your training is landing.
3. Hand it a file
Drop a file onto the Materials panel on the left (or press ＋) and it's stored in a dedicated knowledge base — ◯◯’s materials — converted to OKF so it can be searched and referenced. The material survives even if the original file is moved or deleted, and it follows this agent automatically whenever you bring them into a thread (equipped knowledge).
You can also hand files over via 📎 in the assistant on the right. That way it reads the contents and proposes entries for the judgment rules and knowledge as well. PDF, Word, Excel, Markdown and the like can be handed over as they are.
4. Hand over from a thread
What you settle while talking in a thread can go straight to an agent. Press ⋮ beside the send button → Hand over to an agent and pick who takes it.
| Recipient | What happens |
|---|---|
| Create a new agent | Claude proposes a name, icon and category, and the agent is created on the spot (rename it whenever you like) |
| An existing agent, or your twin | That agent's page opens with the suggestions waiting |
The conversation is read and turned into entries the agent can actually use next time, sorted into judgment rules, knowledge, voice and approach. As with every other route, you decide one by one what to keep.
Today's deadline or a number that only applies to this job is left out — only what will still be true next time comes across. Entries that overlap with what the agent already has arrive as revisions, and ones that clash arrive as contradictions, so you can choose which to keep.
The difference from Hand over to the AI team is doing the work versus remembering it. The team handoff gets the job executed; the agent handoff means you won't have to explain the same thing again.
When it contradicts, and when it fills up
When a proposal clashes with an entry that already exists, you're asked to choose between three options.
| Choice | Result |
|---|---|
| Replace | Swaps the old entry for the new wording |
| Keep both | Keeps both (for when each applies in its own case) |
| Skip | Doesn't take it in |
As a section approaches its cap, you'll be offered a merge that folds similar entries into one. Merging keeps the originals in the History, so you can always trace what was folded into what.
Equipped knowledge
Register knowledge under Equipped knowledge in ⚙ settings and it comes along automatically whenever you bring that role into a thread or add it to an AI team. No more picking the same knowledge out again every time.
For example: equip the reviewer with your Pharmaceutical and Medical Device Act knowledge, and it's referenced automatically every time you ask for a review.
Auto-observe
Switch on Auto-observe in ⚙ settings and, whenever you correct or coach that role in an ordinary thread, it picks up just that moment and saves it as a training proposal. Proposals arrive in the role's assistant marked Observed.
Because it reads what's said in conversation, it does nothing until you enable it. Turn it on per role, as you need it. Even switched on, no proposal is ever applied without you.
Duplicating, disabling, resetting, deleting
| Action | What it does |
|---|---|
| Duplicate | Copies a built-in role so you can rework it as your own. Name, icon and contents are all yours to change |
| Disable | Hides a role you don't use. It also drops out of the team line-up options. Reversible at any time |
| Reset to default | Returns a built-in role to the state it shipped in |
| Delete | Removes the role from the list — built-in roles included. What you trained into it goes too |
Built-in roles can be deleted too. Deleted roles stay at the bottom of the list under Deleted agents, and Restore brings one back at any time — but it comes back fresh: the training does not return. The confirmation tells you how many entries are about to go.
Disabling keeps everything you trained and only hides the role from the team line-up. Delete is for clearing out roles you have no plans to use again.
Name, icon and photo
Click the name at the top of the screen to change it on the spot. The 17 built-in roles can be renamed too. That way “Reviewer (quality)” can be brought into line with whatever your company calls it.
Existing threads, AI teams and loops that use the role carry on working after a rename. For a built-in role, Reset to default brings the original name back.
Click the icon on the left and you get 📷 Choose a photo and 😀 Change icon. A role holds a photo and an icon separately.
| Where | What shows |
|---|---|
| The agent screen and the list on the left | The photo (the icon if there isn't one) |
| Equipped chips, team line-ups, the role tag on a thread | Always the icon |
Photos turn to mush wherever they sit small and side by side, so those places stay as icons.
Handing a role over (sharing and importing)
A role you have grown can be written to a single file and handed to someone else. Share in the header writes it out, once you have looked over what is in it. The person receiving it only has to take it in with ⤓ on the list heading.
Equipped knowledge, history and operational settings such as auto-observe are not included, so it never points at things the other person doesn't have and none of your internal history goes with it. What crosses over is the judgment rules, knowledge, style and approach, plus the role definition.
It looks for anything resembling an API key, a personal folder path, an email address or a phone number, and points at the lines it found. You can share it anyway — but if something looks wrong, close this and fix the entry first. The person receiving it can look over the contents and the same warnings before importing.
An imported role is added under Yours. Nothing is overwritten, even if the name already exists. Your digital twin is you, so it is never written out.
Default model
Default model in ⚙ settings pins the model that role uses. What used to be “per-role model defaults” on the settings screen now lives here, on each role.
History
Open History at the bottom of the screen to see, in order, when something was added and to which section, and what was merged or removed. How the role grew becomes its own record.
Where to go next
- AI team — splitting work across roles
- Knowledge — registering the knowledge that equipped knowledge draws on
- Your digital twin — the layer that learns you
- Skills — how leaving something as a procedure differs