Glossary
Every Surogate term you’ll see in Work mode, in one place.
| Term | Definition |
|---|---|
| AI disclosure | The per-agent EU AI Act (Art. 50) notice telling end users they’re talking to an AI — a web banner, or the first message on messaging channels. Levels: Basic / Enhanced / Full / None. |
| Agent | A deployed AI service with a URL like <slug>.cloud.surogate.ai. Bundles a model, skills, knowledge bases, MCP servers, persona, and channels. |
| CLARIFY | A built-in confirmation card the agent (or Copilot) shows before destructive or expensive operations. You click Yes or No. |
| Allowance (per-user) | The usage budget one end user draws from on one agent: a period grant (subscription tier, trial, or free-mode cap; resets on cycle) plus a top-up bucket (token packs; never expires). |
| Browsing time | Metered wall-clock minutes a buyer’s agent browser may stay open. Distinct from the project’s own web-browsing wallet. |
| Buy page | The hosted storefront at /buy/<agent-slug> where buyers sign in, purchase offers, and manage subscriptions. |
| Buyer | An end user who purchased access to a monetized agent. |
| Channel | The way users reach the agent — Web, Slack, Telegram, Website widget, API. Same agent, different doors. |
| Channel identity | A link between a platform-specific user (a Slack user, a Telegram user) and an internal Surogate user. Lets the same person on Slack and Web share sessions. |
| Copilot | The platform’s built-in AI agent. Press ⌘+/ to open. Runs platform ops on your behalf. |
| Default package | The package applied to users with no active purchase — set from the Free trial / Per-user limit cards on the Monetize tab. |
| Develop mode | The Studio surface (/studio/*) — for builders. Adds Models, Compute, Hub, Datasets, Training, Evaluations. |
| EVENTS tab | The full ordered log of everything that happened in a session: user messages, LLM responses, tool calls, results, decisions. |
| Entitlement | What one buyer holds on one agent: subscription status, balances (usage, browsing, media), and their effective package across the subscription and pack planes. |
| Free trial | A one-time usage grant for new users of a monetized agent, advertised on the buy page. Never refills; a lapsed subscriber gets no second trial. |
Goal (/goal) |
A slash command that defines an outcome. The agent keeps working until an evaluator says satisfied / failed / blocked, or 20 iterations are used. |
| Governance policy | An agent’s enforced guardrails: tool allow/deny, web & browser egress, and AI disclosure — checked before every tool call, on top of the always-on platform floor. |
| Inbox | A per-user queue of items the agent has pushed up for attention — ask_user_question prompts, action requests, completion notices. |
| Knowledge base (KB) | A curated corpus of documents the agent can browse at runtime via kb_list_pages / kb_read_page. Sources → Curator → Wiki. |
Loop (/loop) |
A slash command that schedules a prompt to run on a recurring basis (fixed or dynamic interval). |
| MCP server | An external system the agent can use as a tool — GitHub, Slack, your CRM, internal APIs. Follows the open Model Context Protocol. |
| Media credits | The prepaid media wallet, in dollars of provider spend for image/video generation. Also sellable to buyers on offers. |
| Message (usage unit) | ≈ 1,000 tokens — one question plus the agent’s answer. The unit commerce surfaces display; storage stays token-denominated. |
| Model tier | Basic or Pro — which platform model class a buyer’s package entitles them to. Strongest tier wins across planes. |
| Offer | A plan or pack you sell on one agent: price + usage + browsing time + media credits + a package. |
| Package | What a purchase unlocks beyond usage: capabilities, channels, KBs, connected tools, skills, model tier. Snapshotted at purchase. |
| Persona (SOUL.md) | The agent’s identity, tone, and rules. Up to 20,000 characters. Edited in the CONFIG tab. |
| Platform floor | The always-on, non-configurable governance layer: workspace path containment and path-argument hygiene. No agent policy can relax it. |
| Pricing model | How an agent charges: Free, Subscription, One-time packs, or Both. |
| Quality flags | Per-session tags: 👍 / 👎 (user feedback), policy.denied, harness.crash, saga.compensated, expert.override, expert.endorse. |
| Replay | Fork a session: re-run your user messages against the new agent config to see if the new behaviour is better. |
| Session | One conversation between a user and an agent. Records everything. Survives crashes. |
| Skill | A small markdown file telling the agent how to handle a specific kind of request. Reusable. |
| Tool | A capability the agent can invoke — kb_read_page, web_search, read_file, terminal, browser_navigate, etc. The agent decides which to use. |
| User (per-agent) | An end-user account on a specific agent. Different from your operator account. |
| Vault | Encrypted credential storage for MCP servers. Secrets never reach the agent’s sandbox. |
| Wallet | A prepaid consumable balance (tokens, browsing minutes, compute credits, media credits): plan-granted portion resets monthly, top-ups roll over, operations stop at zero. |
| Work mode | The operator-facing surface (/work/*) — Agents, Sessions, Library. You’re here. |
What’s next
Section titled “What’s next”FAQ for common questions.