Core concepts
Agent
An agent is an identity, plus a focused set of capabilities, plus its own memory. Its working process lives not in the agent itself but in a dedicated orchestration skill (Experiment Design’s, for instance, always runs “multiple candidates → critique → revision”). The platform ships a number of built-in agents. For R&D, for example: Experiment Design, Formulation Development, Ingredient Screening, Data Analysis Expert, Aesthetic Medicine R&D. For marketing: Market Insight, Competitor Intel, Product Proposal. And Office Editor, which works only on the one office file you have open (Word / Excel / PPT in the artifact panel). The Agents tab in the Capability Center lists every agent available to you. An agent is used only when you pick it for a new conversation; the default assistant never calls agents on your behalf.
The skill allowlist is a hard gate: the agent can neither see nor call capabilities that are not on it. This keeps each agent’s behavior predictable instead of every agent doing a little of everything.
Skill
A skill is the smallest unit of capability — one literature search, one sample-size calculation, one INCI compliance check, one report render. Agents orchestrate them per task.
The skill library lives in the Capability Center (sidebar → Capability Center → Skills), listing every skill available on the platform, filterable by source (public / private) and function.
Conversations are bound to an agent
Once a conversation starts, its agent cannot be swapped. This is not an arbitrary limit but a correctness requirement: an agent works out how far along a task is from the files already produced in the workspace, so swapping midway leaves the new agent facing intermediate results it never made, with a different skill allowlist and way of working on top.
To switch agents, start a new conversation; to carry material over, save the artifacts to the asset library and attach them.
Artifact
The files an agent produces in a conversation: reports, tables, images, decks, data files. They are real files you can preview and download, not a paragraph in the chat transcript.
Asset library
A file pool that outlives any single conversation. Files you upload and artifacts agents produce can be stored here, then attached from the input box in any later conversation. This is the recommended way to carry results from one task into the next.
Brand profile
The Brand profile tab of the Assets page holds your brand profile: your brand, product lines, positioning, and tone of voice. The platform adds it automatically to brand-related work (a brand-facing agent, or a brand skill used in chat), so you never have to restate it.
Memory
The platform remembers your research context, recurring terminology, and preferences across conversations, so you don’t have to restate them every time. Review and clear it in Settings → Memory. Memory and the brand profile are different things: memory is what the platform observes about you, while the brand profile holds your brand’s established information.
Thinking tiers
The same agent can work at different tiers: Flash is fastest, Thinking turns on reasoning, and Pro and Ultra spend more reasoning and verification rounds. Higher tiers are slower but produce more complete output. Ultra also includes a three-stage deep-research workflow: research, landscape, and review tasks automatically go through “plan → evidence → report”, with graded evidence and citation checks.
Some agents have their own default tier (Experiment Design defaults to Pro). Once you pick a tier, the platform remembers it.
Toolbox
The Toolbox in the sidebar is a set of tool cards, each with its own workbench: fill in parameters, run, and get results without a conversation. There are 35 cards today, in ten groups:
| Group | Tools |
|---|---|
| Screening & Docking | Molecule–Pocket Binding Score, Protein–Ligand Docking, Metal–Enzyme Binding Prediction, ADMET & Peptide Properties |
| Formulation & Stability | Ingredient Synergy, Formula Compatibility, Composition Check, Phase Check |
| Property Calculators | HLB Calculator, HSP Distance, Mixture Activity Coefficients |
| Sensory & Skin Feel | Taste Prediction, Scent Screening, Skin Feel Prediction |
| Market | Ingredient Trends, Buyer Insight, Creator Rankings, User Interviews, Key Visual Studio, Image to 3D, Infinite Canvas |
| Social Listening | Sentiment Monitor |
| Social Data Collection | Trending Topics, Content Radar |
| Competitor Intel | Competitor Intel, Product Link Parser |
| Office | Office Files (make a deck, write a doc, make a table) |
| Ingredients & Stability | Ingredient & Supplier Library, Industry Frontier |
Computation is built in (in-browser math, the model, external compute services, or read-only data). Results come with charts, can be exported, and are billed in credits: charged on success, never on failure.
To keep calculation results long term, export them or save them to the asset library. User Interviews saves research projects, interview records, and reports: describe the research question → design the discussion guide → conduct interviews → organize records → select interview material and generate an insight report.
Projects
Organize conversations into project folders. A conversation belongs to at most one project. A project can have project instructions (up to 2,000 characters, added to every conversation in the project, including subagents), plan and progress (free text, added alongside the instructions), and a project library (its file list is added with the instructions, and the agent reads entries by id with read_asset). Deleting a project only unlinks its conversations and assets; nothing is deleted.
Scheduled jobs
Tasks that run on a schedule: at the set time, a scheduler starts one agent run, and each run creates a new conversation with a predictable name (job title · MM-DD HH:mm). Results can be viewed inline on the job page. Use it for recurring work such as tracking new research or weekly usage reports. Deleting a job also deletes the conversations it created; the asset library is not affected.
Image understanding
The conversation model is natively multimodal, so you don’t need to worry about which tier can see images:
- Images — the model looks at them itself after upload (it calls the image tool on its own; you do not have to prompt it);
- PDF / PPT / scans — the agent extracts page images and passes them to the image tool for description; for word-for-word transcription there are also free local OCR and text-layer tools.
Vision is an enhancement, not a requirement: if the external vision service is down or times out, everything degrades gracefully and the conversation is never interrupted.
Credits
Accounts are invitation-only and come with an initial credit balance. Conversations and toolbox tools are billed in credits. When the balance reaches zero, running tasks stop and new ones are blocked (you can still sign in); once an admin adds credits, everything resumes automatically.
Orphic