Features

Knowledge Agents

Every organisation gets four built-in domain experts โ€” Finance, HR, Legal and Procurement โ€” each answering from its own knowledge collection of the organisation's policies, with a citation on every claim and an honest refusal when the answer is not there. They live in the org ring, are visible from every workspace, and are candidates for the Orchestrator, so "what is the daily meal allowance on domestic travel?" reaches the Finance Agent without anyone naming it.

What a person sees

Ask the default chat, or mention the agent directly:

@finance-agent what is the daily meal allowance on domestic travel?

The Finance Agent searches the finance-policies collection, finds the travel policy, and answers with the figure as the document states it โ€” the domestic allowance, the departure-day rule, the receipts-or-allowance choice โ€” each sentence carrying a [chunk:โ€ฆ] citation that opens the passage it came from. Ask the same agent about parental leave and it says the finance corpus does not cover it and points you to the People team: the HR policy exists in the organisation, but it is the HR Agent's corpus, not this one's.

The four experts

AgentHandleCollectionSample corpus covers
Finance Agent@finance-agentfinance-policiesExpenses and reimbursement, travel allowances, corporate cards, spending and approval limits, budgeting and the fiscal calendar
HR Agent@hr-agenthr-policiesLeave and time off, remote and hybrid working, performance reviews, onboarding and probation, the code of conduct
Legal Agent@legal-agentlegal-policiesContract review and signing authority, NDAs, data protection and privacy, intellectual property and records retention
Procurement Agent@procurement-agentprocurement-policiesPurchase orders and competitive quotes, vendor onboarding and due diligence, software and SaaS purchasing, the supplier code of conduct
The seeded documents are samples Each collection starts with policies of a fictional Trellaris organisation, written so that concrete questions have concrete, citable answers. Every document says so in its header. Replace them with your own โ€” upload into the same collection, or attach a synced source โ€” and the agent needs no change: it reads whatever the collection holds.

How it is built

There is nothing new in the runtime. A Knowledge Agent is an ordinary versioned markdown agent whose frontmatter names one collection, and the platform's existing knowledge path does the rest โ€” search_knowledge switches on, the citation contract is injected, and retrieval respects the ring the collection is in.

---
name: finance-agent
model: default
tools:
  - search_knowledge
  - calculate
  - current_time
knowledge:
  - finance-policies
---
  1. Seeded per organisation

    On sign-in the platform makes sure the four agents and their four collections exist in the org ring โ€” one copy each, visible from every workspace. The corpus itself is ingested by a worker in the background, so a sign-in never waits on it.
  2. Searched, then answered

    The definition insists on searching before answering, citing every claim, and saying plainly when the corpus has nothing โ€” never filling a gap with general knowledge presented as your policy.
  3. Routed like any specialist

    The agents are routable Orchestrator candidates. Their tool list, as the router sees it, includes search_knowledge โ€” so a plan step that needs a policy looked up is sent to an agent that can, not filtered away.
  4. Yours to edit

    Each agent is a normal registry entry with versions and rollback. Change the boundaries, add a tool, rename the collection โ€” the seeder heals only copies it wrote itself and keeps your edits.

Where they show up

Chat

Mention @hr-agent, pick the agent, or just ask โ€” the fast path names the owner when one specialist can answer the whole thing.

Orchestration

A multi-step goal that includes "check what the policy says" gets that step routed to the right expert, and its answer cited back in the synthesis.

Knowledge

The four collections appear on the Knowledge page like any other org collection; add documents or a synced source there.

Adding your own domain

The pattern is the whole feature: create an org-ring collection, upload the documents, and author an agent whose knowledge: names it, using the definition above as the template. Promote it to the org ring and it is routed exactly like the built-ins. A domain that needs to act โ€” file a ticket, raise a purchase order โ€” is a different kind of agent; see Connectors.