Understanding AI Agents
A deeper look at how Agency Hero's AI agents actually work — the skill architecture, how workspace context is assembled, how agents take action, and what they can and can't do autonomously.
If you’ve read Getting Started with Chat, you already know what the agent can do. This article explains how it actually works — the architecture behind the responses, the way context is assembled, and the rules that govern when the agent acts independently versus when it pauses to ask you.
This is the foundation for everything else in this section. Understanding the agent’s design will help you use it more effectively and set the right expectations for what it does autonomously versus what it hands back to you.
The mental model: an agent is not a chatbot
Most AI chat tools are stateless — every message is a fresh start, and the model has no idea who you are or what you’re working on. Agency Hero’s agent is fundamentally different.
Think of it as a context-aware operator rather than a chatbot. When you send a message, the agent doesn’t just look up a likely response. It:
- Understands your intent — classifying what kind of task this is (search, action, research, creation)
- Loads the right skill — a specialized set of instructions, tools, and permissions for that task type
- Assembles your workspace context — pulling in the relevant data from your meetings, decisions, topics, and tasks
- Takes action or answers — either responding with synthesized information or executing a tool call
- Checks permissions — ensuring every action is within what you and your workspace allow
Each of those steps is worth understanding in detail.
Skills: how the agent adapts to what you need
The agent isn’t one fixed thing. It’s a runtime that loads a skill — a focused module that defines how to behave for a specific type of work.
A skill specifies:
- The role the agent adopts for this task (e.g., “strategic account analyst” for meeting prep, “research assistant” for in-depth investigations)
- Which tools it can access (e.g., task creation, memo writing, web search)
- What guardrails apply (e.g., require approval before creating tasks, never write to external systems without confirmation)
- How to structure the context it needs (e.g., meeting prep needs recent meetings and open items; research needs search indices)
Skills are selected automatically based on what you ask. The agent reads your message, classifies the intent, and routes to the right skill — this is called skill routing, and it happens before any context is loaded or tools are invoked.
System skills in Agency Hero
Agency Hero ships with a set of built-in system skills:
| Skill | What it's for |
|---|---|
| **General Assistant** | The default. Handles search, Q&A, quick tasks, and anything that doesn't match a more specific skill |
| **Meeting Prep** | Deep preparation for upcoming meetings — surfaces context, open items, risks, and talking points |
| **Task Manager** | Creates, updates, searches, and organizes tasks via natural language |
| **Research Assistant** | Multi-step investigations that require iterative searching, web lookups, and synthesis across sources |
| **Doc Author** | Writes, edits, and improves documentation articles grounded in workspace knowledge |
Additional skills are added as Agency Hero’s capabilities expand.
Invoking a skill manually
When you want to direct the agent to a specific mode without relying on it to infer the right skill, use @ followed by the skill’s slug in the chat input:
@meeting-prep— loads the Meeting Prep skill@research-assistant— switches to Research Assistant for deep investigations@doc-author— activates the Doc Author skill for writing and editing@task-manager— explicitly routes to task management
Manual invocation is useful when the agent might otherwise generalize — for example, if you want a thorough research pass rather than a quick search, opening with @research-assistant ensures the right behavior from the start.
→ For a full explanation of how routing decisions are made and when to invoke skills manually, see How Skill Routing Works.
How the agent knows your workspace
The agent doesn’t treat your workspace as a monolith. Agency Hero stores different types of information in specialized backends, and the agent knows which one to query for each type of question.
The four knowledge sources
| Source | What's in it | Example questions |
|---|---|---|
| **Topics** | Recurring themes and workstreams extracted across meetings | *"What topics are we tracking?"* / *"What's the status on the API migration thread?"* |
| **Intelligence** | Structured items extracted from meetings: decisions, risks, open questions, action items | *"What decisions have we made about pricing?"* / *"Show me all open risks"* |
| **Knowledge** | Uploaded and generated files — SOWs, proposals, memos, documents | *"Find the Q3 proposal"* / *"What does the SOW say about scope?"* |
| **Meeting summaries** | AI-generated summaries and transcripts of past meetings | *"What was discussed in Monday's standup?"* / *"Summarize last week's Acme meetings"* |
For broad questions — “What do we know about the DevOps workstream?” — the agent searches all four sources in parallel and synthesizes the results. For targeted questions — “What decisions have we made about pricing?” — it goes directly to the intelligence source.
The context pack
Before responding to any message, the agent assembles a context pack: a pre-loaded snapshot of your workspace’s current state. It includes things like open tasks, tracked topics, recent activity, key stakeholders, and pending items.
The context pack is what allows the agent to answer questions like “What’s my week look like?” or “What’s still open from the kickoff?” instantly — without running a full search. Think of it as the agent’s briefing document for your workspace, automatically prepared before every conversation.
Workspace boundaries
The agent is always scoped to your current workspace by default. If you’re in the Acme Corp workspace, the agent queries Acme Corp data and creates Acme Corp tasks. It won’t blend in data from another client unless you explicitly ask.
For cross-workspace queries — “Show me tasks across all my workspaces” or “Which clients have no meetings this week?” — the agent fans out to every workspace you have access to, runs scoped searches in parallel, and attributes results by workspace. Data from one client is never surfaced in another client’s context.
How the agent takes action
The agent doesn’t just answer questions — it can take actions: creating tasks, logging decisions, generating memos, updating records. Understanding how this works helps you know what to expect.
Tools
Each skill comes with a set of tools — discrete capabilities the agent can invoke to accomplish work. Tools are specific and scoped:
meetings.search_summaries— search meeting transcript summariestasks.create— propose a new task for your reviewintelligence.create— log a decision, risk, or open questionmemos.create— draft and save a memotopics.search— search cross-meeting topic threads
When you ask the agent to do something, it selects the appropriate tool, prepares the parameters based on your request and workspace context, and either executes it or presents it for your approval — depending on the action’s risk level.
The risk tier system
Every tool action in Agency Hero is assigned a risk tier that determines whether it requires your approval before executing:
| Tier | Type of action | Approval required? |
|---|---|---|
| **0** | Read-only — searching, listing, looking things up | No — happens immediately |
| **1** | Low-risk writes — reversible, low-impact internal changes | No — executes automatically |
| **2** | High-risk writes — creates, deletes, external actions | Yes — pauses for your approval |
For Tier 2 actions, the agent presents an approval card: a preview of exactly what it’s about to do, which fields it will set, and what the result will look like. You can review, edit any field, approve, or cancel.
Examples of Tier 2 actions that trigger an approval card:
- Creating a task (you review the title, description, assignee, and due date before it’s saved)
- Logging a decision or risk to the intelligence registry
- Sending an email or posting to an external channel
- Updating CRM data
This design means you’re always in control of what gets written. The agent never creates, deletes, or sends anything without your explicit sign-off.
→ For a detailed explanation of the approval system and how workspace admins configure tool permissions, see Tool Approvals and Permissions.
Subagents
For complex, multi-step tasks — especially ones that require searching across multiple workspaces simultaneously — the agent can spawn subagents: isolated workers that handle a specific piece of the problem in parallel.
For example, when you ask “Give me a status update across all my client workspaces”, rather than querying each workspace sequentially, the orchestrating agent spins up one subagent per workspace. Each subagent searches its assigned workspace, returns its findings, and the orchestrator synthesizes everything into a single response.
Subagent behavior:
- Each subagent has its own isolated context so complex tasks don’t pollute your main conversation
- Subagents inherit the exact same permissions as the parent agent — they can never escalate beyond what you’re allowed to do
- The Task tool (used for spawning additional subagents) is explicitly disabled for subagents themselves, preventing uncontrolled recursion
- Results are aggregated by the orchestrator and presented as a unified answer
In practice, you won’t notice subagents working — the experience is just a faster, more thorough answer for cross-workspace or multi-step requests.
What the agent can and can’t do on its own
Understanding this boundary is important for building accurate expectations.
What the agent does autonomously
- Searches and retrieves information from all four knowledge sources
- Loads context, synthesizes results, and generates answers
- Routes to the right skill based on your message
- Spawns subagents for parallel work
- Executes Tier 0 and Tier 1 tool actions without interrupting you
What always requires your approval
- Creating tasks, memos, or intelligence items (Tier 2 writes)
- Sending anything to an external party — emails, Slack messages, CRM updates
- Any action the workspace admin has configured as approval-required
- Deleting or modifying existing records
What the agent won’t do
- Invent information — if it doesn’t find an answer in your workspace data, it says so rather than generating a plausible-sounding response
- Cross workspace boundaries uninvited — it stays scoped to your current workspace unless you ask it to fan out
- Access data you don’t have permission to see — workspace permissions are enforced at every step
- Take action on behalf of another user — it always acts as you, with your permissions
These aren’t limitations to work around — they’re design choices. When you’re managing multiple client relationships, you want an assistant that respects those boundaries by default and asks before it acts.
What’s next
You now have the foundation: how skills are selected, how context is assembled, how actions work, and where the agent defers to you. The rest of this section goes deeper on specific capabilities and workflows.
Immediate next steps:
- How Skill Routing Works → — A detailed look at how the agent routes your message to the right skill, when routing works automatically, and when to override it with
@skill-name. - Creating and Managing Tasks via Chat → — The complete guide to task management via conversation: phrasing requests, the approval flow, available task fields, and how meeting action items become trackable tasks.
Later in this section, you’ll find guides on asking effective questions, searching and retrieving information, how approvals work in detail, managing chat threads, and advanced capabilities like background research and cross-workspace workflows.
Related articles
More resources to help you go deeper.