For the complete documentation index, see llms.txt. This page is also available as Markdown.

Agents

The ten Clarity agents, the surfaces they lead, and how they divide the work.

Clarity runs one agent kernel under ten roster profiles. A profile sets the agent’s individual instructions, tools, pinned skills, budget and model. Its instructions are editable as one complete prompt. Every profile is configurable per organization in the admin panel; see In the admin panel.

Surfaces, their lead agents, the desks they delegate to, and the evidence and check agents

Roster

Agent
Job
Leads
Delegates to
Steps / deadline

cio

Owns the house view. Frames the question, delegates, reconciles the board, sets conviction and dated tests, writes. The only agent that can propose house theses.

Market update, chat

Every desk, associate and QA

24 / 20 min

macro-strategist

Regimes and shock transmission; a mechanism on every causal link

—

research-associate

16 / 8 min

sector-strategist

Industries and supply chains, with named companies per layer

—

research-associate

16 / 8 min

equity-analyst

One listed name: fundamentals, catalysts, valuation, place in the house themes

Company read

research-associate

16 / 8 min

portfolio-strategist

The book against the house view: exposure, drivers, what changed, what to review

Portfolio card

risk-officer, equity-analyst, associate, QA

20 / 12 min

risk-officer

Concentration, correlation, drawdown, what breaks the book

—

research-associate

16 / 8 min

private-markets-analyst

Underwrites one private company from the holder's documents and independent reporting. Never prices the stake or recommends an action.

Private investment update

—

20 / 10 min

client-advisor

Fit of the house view and the book for one person; proposes investor theses; never gives execution instructions

Investment ideas

—

18 / 10 min

research-associate

Maps the graph and returns evidence, not opinion

—

—

12 / 5 min

research-qa

Checks a draft sentence by sentence against its citations

—

—

8 / 4 min

How work is divided

  1. One lead per surface. The lead owns the answer from question to submission.

  2. Desks on card surfaces only. The market update, company read and portfolio card delegate. Chat, ideas and the private investment update do not.

  3. One level deep. A desk may use the associate for evidence but never delegates further.

  4. The board is the channel. Desks post findings with evidence ids to a shared board. The lead reconciles contradictions and alone writes the answer.

  5. Shared reads flow down. Company reads cite the current market update. The portfolio card cites both.

What every answer follows

  • Cite or omit. A fact without evidence is left out, not narrated as a gap.

  • Newest fact wins, and the answer says what it replaced.

  • Plain language. A number comes with its comparison: against what, over what period.

  • Mechanism over mood. A causal claim names how the driver reaches the effect.

  • Dated tests. Each view states what would confirm or break it, and when to check.

Models

All ten profiles default to GPT-5.6 Luna through OpenRouter, with an optional reasoning effort. An organization can pin another model per agent, or fall back to its own model route. The harness re-sends the whole run each step, so any model must have cached-input pricing; see Agent harness.

Last updated