Skip to main content

Agent aggregate

RaybotAgent is the root of the model. Everything else in the agent module hangs off it.

export interface RaybotAgent {
  identity: AgentIdentity;
  configuration: AgentConfiguration;
  settings: AgentSettings;
  channels: PublishChannel[];
  confidence: ConfidenceLevel;
}

The four members map one-to-one onto the four things a maker actually manipulates in Coral, in the order the runbook manipulates them:

MemberWhere it is editedRunbook phase
identityMinted by the platform, then read-onlyPhase 1
configurationBuild canvas and the right railPhase 1, Phase 2, Phase 3
settingsThe four-tab Agent settings dialogPhase 4
channelsCustomize publish channelsPhase 5, Phase 6

Why the split exists​

The obvious alternative is one flat agent object. The split is deliberate, and it comes from a property of the product: the three groups have different mutability rules.

identity is write-once. configuration is freely editable and takes effect on the draft immediately. settings is freely editable but some of its fields are projections of identity and are therefore read-only. channels is not editable at all in the normal sense - a channel is enabled, and whether that enablement worked is discovered later, out of band.

Collapsing those into one object would make it impossible to express "this field cannot be changed" without annotating every field individually.

Build configuration​

export interface AgentConfiguration {
  model: string;
  instructions?: string;
  icon?: AgentIcon;
  knowledge: KnowledgeSource[];
  tools: AgentTool[];
  skills: AgentSkill[];
  connectedAgents: ConnectedAgent[];
  microsoftIQ: MicrosoftIQBinding[];
  memory: AgentMemory;
}

The five array members are exactly the five add-buttons on the right rail, plus Microsoft IQ. They are arrays rather than optional singletons because the rail permits several of each - even where Raybot only ever used one.

model is a plain string, not a union. Coral's model picker is server-driven and the list changed between the build and the re-audit; freezing it into a union would encode a snapshot as if it were a contract. The value observed on Raybot is Claude Sonnet 4.6.

instructions is optional because an agent can be saved with none. Raybot has instructions; the model does not require them.

The canonical instance​

RAYBOT is a fully populated RaybotAgent and doubles as the reference fixture for the package.

import { RAYBOT } from '@raybot/object-model';

RAYBOT.identity.agentId;        // '4de4ada9-ad61-4c22-a0d0-a0bd1c189f4e'
RAYBOT.configuration.model;     // 'Claude Sonnet 4.6'
RAYBOT.configuration.knowledge; // one publicWebsite source at https://docs.ray.io
RAYBOT.configuration.tools;     // [] - surveyed, never attached
RAYBOT.channels.length;         // 2

Two members are empty arrays, and that is a finding rather than an omission. Raybot surveyed all four tool catalogs in Phase 2 and attached nothing; it opened the connected-agent picker and found it empty. Both facts are recorded as empty arrays with the explanation living in the runbook phase, because an empty array is the honest representation of "we looked and took nothing".

Operators​

RAYBOT_OPERATORS records who touched the agent and in what capacity. Two identities were involved, and conflating them would make the provenance unreadable.

export const RAYBOT_OPERATORS = [
  { upn: 'darbot@timelarp.com',  role: 'build',    note: 'Built the agent end to end through the Coral UI.' },
  { upn: 'dayour@microsoft.com', role: 're-audit', note: 'Ran the 2026-06-17 re-audit. Copilot-DYdev25, bbf34eea-8ed1-e502-a74c-6a7b21a8b002.' },
] as const;

See Identity for the full identity crosswalk.

Aggregate confidence​

RaybotAgent.confidence is CONFIRMED, and that claim is checkable rather than asserted: every Evidenced member below it is also CONFIRMED, so weakestConfidence() over the tree returns CONFIRMED.

import { weakestConfidence } from '@raybot/object-model';

weakestConfidence([
  ...RAYBOT.configuration.knowledge.map(k => k.confidence),
  ...RAYBOT.configuration.skills.map(s => s.confidence),
  ...RAYBOT.configuration.microsoftIQ.map(m => m.confidence),
  ...RAYBOT.channels.map(c => c.confidence),
]); // 'CONFIRMED'

If any component were downgraded, the aggregate claim would have to move with it. See Confidence model.

  • Identity - the write-once half and the first-save minting behaviour
  • Settings - the four tabs
  • Publishing - PublishChannel and the two-field enablement model