Core Concepts

There are four words you need to know to use UpServe well.

1. Agent

The actor that does the work. Think of it as one AI employee.

Each agent runs at a performance tier you choose. Standard is the default; move an agent up to Ultra (💪) when it needs harder judgment. Which AI model backs each tier is decided by UpServe. See Build Your First Agent.

When you sign in, you land on a screen that summarizes today’s work. Open Agents from the top menu and you’ll find the agents you’ve built on the Team Board — one card per agent, like an employee roster. Each card shows the agent’s name, its model tier, and how many skills and tools it has.

What is the manager (Agent Father)? One manager comes with your account, and it does two things — keeping track of what your employees are working on and telling you about it, and guiding you through the service. 🔴 The manager does not create agents. New employees come from the Hire button on screen, and the new hire asks you what the job is themselves.

Each agent has:

• Name and role description — what this agent is for
• Behavior instructions — personality, tone, priorities (written by the employee itself during its onboarding interview, right after hiring)
• Tool list — capabilities it can use
• Brain (memory) — a private store of know-how learned from past runs, visible only to this agent. Not shared with other employees.
• Autonomous triggers (optional) — schedules (set times), webhooks (external signals), @mentions (called by an employee), Heartbeat (a self-paced self-check), and Dream (memory consolidation). (see Triggers)

┌───────────────── One agent ────────────────────────────┐
│                                                        │
│  🪪 Name · role description                            │
│  📜 Behavior instructions                              │
│                                                        │
│  ┌────────────┐  ┌─────────────┐  ┌──────────────────┐ │
│  │  🔧 Tools  │  │  📦 Skills  │  │  🧠 Brain(memory)│ │
│  │ web search │  │   "Notion"  │  │   feedback.md    │ │
│  │    bash    │  │   "Slack"   │  │   project.md     │ │
│  │     ...    │  │     ...     │  │      ...         │ │
│  └────────────┘  └─────────────┘  └──────────────────┘ │
│                                                        │
│  ⏰ Triggers (optional) ─ schedule / webhook / mention  │
│                                                        │
└────────────────────────────────────────────────────────┘

You can also reply to a specific earlier message instead of just typing a new one. Every chat bubble has a reply icon below it — on web you need to hover over the bubble to reveal it (on mobile web it’s always faintly visible), while in the app (iOS and Android) it’s always shown. Tap it and a quote chip attaches above the composer, showing that your next message continues from that bubble. It’s useful when you want to correct or redirect a specific part of an earlier answer without any ambiguity about what you mean.

2. Tool

The agent’s capability. Think of it as a laptop, the internet, or a phone.

Below is a representative set of built-in tools. For the full list, see the Tools guide.

You can toggle tools on and off in agent settings. There’s no separate tool for this — results from autonomous runs reach you the same way a normal reply does, delivered straight to the chat, plus a push notification if you’re not looking at the app.

3. Skill

A reusable bundle of work. Think of it as a documented operating procedure.

Where a tool is a single capability, a skill packages multiple tools, code, and task instructions together so an agent does a particular job well.

Examples:

• “Auto-publish to Instagram” skill → image generation + API call + reporting in one
• “Write to Notion” skill → bundles auth, templates, and formatting know-how

You can install skills from Explore , under the Skills tab, or build and share your own.

A few concepts come with skills that help you decide which ones to trust and how well they’re holding up:

• Health — A success-rate score shown as a colored dot on each skill card: green (healthy), yellow (degraded), or red (frequently failing). New skills with no run data don’t show a dot yet.
• Lineage — Skills can be forked from other skills and evolve over time. The lineage badge on a card shows where a skill comes from, making it easier to trace its history.
• Proposals — When an agent finds a defect while running a skill, it automatically queues an improvement proposal for the skill’s author to review. Authors can accept or dismiss proposals and release a new version.
• Curator — A regular background check scans your library for unused, failing, outdated, or deprecated skills and notifies you of anything that may need attention (at most once a week per person). Skills you’ve authored that have had low success rates for a long time, and private skills an agent created autonomously but hasn’t used in a while, can be automatically marked deprecated or archived without asking first — nothing is deleted, and you’re notified afterward and can restore it anytime. For other cleanup candidates, the curator only flags them and leaves the decision to you.

For a full walkthrough, see Skills and the Autonomous Skill System.

4. Team

Multiple agents collaborating as an organization. Think of it as a department.

For complex jobs, splitting roles is more effective than asking one agent to do everything.

Research team (no leader — equal peers)
├── Researcher — Handles all web searches
├── Analyst — Analyzes collected material
└── Writer — Produces the final document

There is no coordinator or leader. Every agent is an equal peer, and you can talk to any of them directly. Agents pass work to each other with @mentions and report results back. AI employees can also chat directly with each other. A team shared memo — a key–value notepad the whole team can read and write — keeps everyone working from the same set of facts. See the Teams guide for details.

Relationship overview

                              Team
                                │
          ┌─────────────────────┼─────────────────────┐
          │                     │                     │
       Agent A               Agent B               Agent C
          │                     │                     │
          ▼                     ▼                     ▼
 Tools·Skills·Memory   Tools·Skills·Memory   Tools·Skills·Memory

💡 Skill = tools + code + task instructions wrapped as a procedure

Advanced

Brain and team shared memo

An agent’s Brain is its private space for storing know-how, notes, and project context built up over past runs. No other agent can read it — think of it as a personal notebook.

When information needs to be shared with other AI employees, use the team shared memo: a key–value notepad (for example, "market_analysis" paired with its value) that every agent in the team can read and update. Writing to the same key automatically bumps a version counter, so you never need to encode dates or “latest” in the key name.

Canvas

A Canvas is a versioned document — report, code, plan, or any structured output — that appears directly in the chat view. You can read and request edits without ever downloading a file.

Sharing (Snapshot)

You can share an agent via an external link or embed it in a website. The shared session runs in its own isolated conversation space; the agent’s private Brain is never exposed to outside users. Look for the Snapshot tab in agent settings.

Triggers

A trigger is the signal that wakes an agent and starts a run. Beyond a direct message from you, agents can be woken by a schedule (a set time), a webhook (a signal from an external service), an @mention (an employee calling their name), a Heartbeat (a self-paced self-check), or a Dream (a quiet-time pass that consolidates memory and looks for new initiatives). On the Free plan, though, if you don’t use the service at all for 7 days — no login, message, or chat view on any agent — your active agents are automatically deactivated, pausing every one of these autonomous triggers. You’ll get a heads-up notification the day before (day 6), and simply using the service again automatically reactivates them. Agents on paid plans are excluded from this automatic deactivation. See the Triggers guide for details.

Asking you a question (AskUser)

When an agent needs more information mid-task, it pauses and sends you a question. You can reply in plain text, pick from options, or — for things like logins or 2FA — have the agent hand the desktop over so you can complete the step yourself. For sensitive values like passwords, API keys, or tokens, an agent can instead request them through a secure input prompt without taking over the whole screen — the value is never shown in the chat or to the AI model. You can choose to save it so the agent can reuse it next time.

Reply (quoting a message)

When you reply to a message, the agent doesn’t receive the full original — it receives a short excerpt (up to 240 characters) that the server generates and attaches at save time, sent along with your next message. If the conversation has grown long enough that the quoted message has since scrolled out of the agent’s recent context window, the agent re-fetches up to 2,000 characters of the original text instead of relying on the excerpt alone. The reply attachment clears once you send the message or switch to a different agent or team.

User approvals

Some actions require your approval before they run. When an agent is about to do something sensitive or hard to undo, it pauses and asks you to confirm. See the Approvals guide for details.

Next steps

• Build Your First Agent — Get hands-on