Skip to main content
Knowledge items are how you give agents durable context — company background, style guides, SOPs, research, files, and standing rules — without repeating yourself every conversation. The knowledge system replaces the old documents and assets model with a single, unified library.

​
How it works

Every knowledge item has a kind (what it is) and a scope (who can see it):

Kind

Page — rich-text documents (company briefs, style guides, priorities). Skill — structured procedures and SOPs. File — uploaded files (images, PDFs, spreadsheets, audio, video). Source — externally synced content from connected integrations.

Scope

Workspace — visible to all agents in the workspace. Pin it to include in every agent’s boot context. Agent — visible only to specific agents, assigned via the agent junction.

​
Giving agents boot context

Boot context is what an agent reads before doing any work. Two mechanisms control it:
MechanismWhat it does
Workspace scope + pinnedA pinned workspace-scoped item is injected into every agent’s startup context.
Agent scope + agent assignmentAn agent-scoped item assigned to specific agents is injected only for those agents.
1

Create a knowledge item

Go to Knowledge in the left sidebar and create a new page, skill, or upload a file.
2

Set scope and visibility

For all agents: set scope to Workspace and toggle Pinned on.For specific agents: set scope to Agent and assign the agents who should receive it from the agent picker.
3

Agent reads it automatically

The next time an agent wakes up (via a message, a task, or a routine), it reads the item as part of its startup context. No restart required — changes take effect on the next session.

​
Organizing knowledge

Use folders to organize items by topic. Folders support nesting and custom icons. Use tags for cross-cutting labels (e.g. “q2-2026”, “onboarding”, “legal”). Tags are separate from scope — they’re for your own organization, not for controlling agent access. Use pinned to surface important workspace items in boot context without having to remember which folder they’re in.

​
Keeping knowledge effective

Large boot context slows agents down and reduces their effectiveness. Context has a limit — the more you pack in, the less room the agent has to think about the actual task.
Keep each item focused. One item per topic is better than one giant document with everything. An agent reads all items in its boot context, so you can have multiple short ones. Keep them current. An agent reading stale knowledge (“Q1 priorities” in Q3) is worse than reading nothing — it acts on outdated information. Review pinned items quarterly or when your situation changes significantly. Signs knowledge is stale:
  • The agent references old project names or priorities
  • It uses outdated context in responses
  • It contradicts what you’ve told it recently in conversation
To update: edit the item directly — changes take effect on the next session.

​
Good boot context examples

Who you are, what the company does, who the customers are, your stage, your voice. 3-5 paragraphs maximum. Every agent benefits from this context.
Tone, vocabulary, formatting preferences, what to avoid. Examples of good and bad outputs help more than abstract rules. Pin it workspace-wide if all agents should follow it, or scope it to specific writer agents.
What matters right now — 3-5 bullet points. Update at the start of each quarter. Helps agents prioritize when they have discretion.
Hard rules that should never be broken: “never commit to pricing”, “always CC [name] on outbound”, “don’t share confidential deal terms”. Short list, no ambiguity.
Step-by-step workflows for recurring tasks — e.g., how to research a new contact, how to structure a newsletter draft, how to handle inbound inquiries. Role-specific, so assign to the relevant agents only.

​
Agent-created skills

Agents don’t just consume knowledge — they create it. When an agent discovers a reusable method during work, it can autonomously create a skill scoped to itself. This happens via the hq_skill_upsert.py skill. When agents create skills:
  • After performing the same sequence 3+ times successfully
  • After a breakthrough discovery (figuring out a non-obvious method through trial and error)
  • When the user gives a reusable instruction that isn’t already documented
What you see:
  • The agent detail page shows a Skills section with the agent’s self-created procedures
  • A blue dot indicates recent edits (within 7 days), with the one-line reason displayed below
  • The knowledge editor’s History panel shows who made each edit (agent name + emoji vs “You”)
  • The dashboard fleet card shows how many skills were updated this week
No approval gate. Agents write directly. You observe via the audit trail and can edit, refine, or delete skills at any time through the standard knowledge editor. If you correct a skill, the agent will respect your edits on subsequent sessions.
Knowledge items can be linked to tasks, routines, collection records, and agents using entity links. Links are bidirectional references — they show up on both the owner and the target. When an agent claims a task, it receives all linked entities (knowledge items, contacts, organizations, collection records) as context for that task. This replaces the old task attachments system with a universal linking model. You can also link external URLs. URL links show up alongside entity links but open in a new tab instead of navigating within HQ.