Starlight Intelligence Lab
Starlight field guide · Foundation 01Public design system · no runtime authority

Profiles · swarms · workflows · trust · evolution

Design the team.Keep the key.

A reusable foundation for building agent profiles, composing swarms, tracing work, managing boundaries, and evolving capability without losing human authority.

  • 10canonical swarms
  • 50mapped agents
  • 10system contracts
  • 50next visual briefs
Reference control circuitEvery path ends in proof or a human decision.
Human sovereignMission · veto · credentials · release
  1. 01
    MissionHuman-owned objective
  2. 02
    ConductorRoute and synthesize
  3. 03
    SpecialistsProduce bounded artifacts
  4. 04
    CriticFind counterexamples
  5. 05
    VerifierAccept or hold
  6. 06
    ReceiptBind proof to the run
  7. 07
    EvolutionChange with lineage

Active council target: three to five roles. Direct consensus ceiling: seven.

Foundational control plane

Roles before personalities.

These are compositional responsibilities, not seven new branded agents. Give each responsibility one visible owner; let identities and models change underneath the contract.

  1. H

    Human sovereign

    Authority

    OwnsMission, values, veto, credentials, irreversible decisions, and final accountability.

    Must notBecome an ornamental approval step after the system has already acted.

  2. C

    Conductor

    Coordination

    OwnsDecomposition, capability routing, sequencing, conflict surfacing, and synthesis.

    Must notAccumulate every specialist capability or silently expand the mission.

  3. S

    Scout

    Discovery

    OwnsSignals, sources, assumptions, unknowns, and opportunity framing.

    Must notTurn unverified discovery into a production claim or decision.

  4. B

    Builder

    Execution

    OwnsA bounded implementation surface and its working artifact.

    Must notReview its own high-risk work as the independent release authority.

  5. K

    Critic

    Adversarial quality

    OwnsCounterexamples, quality pressure, assumption testing, and revision requests.

    Must notBlock forever without a criterion, evidence request, or stop rule.

  6. T

    Steward

    Continuity

    OwnsMemory hygiene, budgets, provenance, version lineage, and resource boundaries.

    Must notTreat storage, context, or compute as unlimited and invisible.

  7. V

    Verifier

    Release evidence

    OwnsIndependent acceptance, failed-gate visibility, receipt integrity, and rollback readiness.

    Must notConvert missing, skipped, or structural-only evidence into a pass.

The primitive spine

Ten contracts make a swarm operable.

If a system cannot show these objects, it may still demonstrate intelligence—but it cannot yet show ownership, safety, continuity, or release truth.

  1. 01

    Mission brief

    starlight.mission_brief.v1

    What outcome matters, who owns it, and what is out of scope?

    A bounded objective, acceptance tests, budget, deadline, and human owner.

  2. 02

    Agent profile

    agent-card.v1

    Who is this agent, what can it do, and where must it stop?

    Identity, purpose, capabilities, non-capabilities, tools, memory scope, gates, and evals.

  3. 03

    Skill contract

    SKILL.md + manifest

    What portable behavior can the agent reliably activate?

    Versioned inputs, outputs, dependencies, examples, tests, and failure behavior.

  4. 04

    Swarm topology

    starlight.swarm_profile.v1

    Which roles collaborate, in what shape, and under whose authority?

    A small active council, routing rules, write scopes, verifier, and escalation path.

  5. 05

    Task packet

    starlight.task_packet.v1

    What exact unit of work crosses the boundary?

    Objective, context references, inputs, expected output, constraints, acceptance, and expiry.

  6. 06

    Memory boundary

    starlight.memory_policy.v1

    What may be remembered, retrieved, shared, or forgotten?

    Named scopes, retention, provenance, privacy class, and a public/private transfer rule.

  7. 07

    Tool policy

    starlight.tool_policy.v1

    Which actions are read-only, reversible, metered, or human-gated?

    Allow, deny, approval, cost, credential, side-effect, and rollback declarations.

  8. 08

    Handoff packet

    starlight.handoff.v1

    How does useful context move without dragging the whole conversation?

    Completed work, decisions, evidence, unresolved risks, next action, and artifact references.

  9. 09

    Run receipt

    starlight.run_receipt.v1

    What happened, what changed, and what proves it?

    Inputs, actor, tools, outputs, checks, cost, duration, verdict, and rollback evidence.

  10. 10

    Evolution release

    starlight.evolution_release.v1

    Why is the next version more capable—and still trustworthy?

    Version delta, migrations, compatibility, eval comparison, lineage, and retirement notes.

Coordination pattern library

Choose a shape that matches the work.

“More agents” is not a topology. Start with the smallest pattern that preserves ownership, context, recovery, and independent verification.

01Conductor + specialistsOne routing authority · two to five bounded specialists · one verifier
MissionConductorScoutBuilderVerifier
Use when
The mission has dependencies, a clear owner, and several distinct capabilities.
Avoid when
The conductor becomes a god agent or every decision must pass through one slow bottleneck.
Handoff
Mission packet → specialist artifact → synthesis packet → independent verdict.
02PipelineResearch → design → build → verify → release
ResearchDesignBuildVerifyRelease
Use when
Stages have stable contracts and later work genuinely depends on earlier output.
Avoid when
A failure at one stage invalidates everything but no checkpoint or fallback exists.
Handoff
Each stage emits one versioned artifact plus a compact context envelope.
03Parallel councilIndependent perspectives → conflict map → weighted synthesis
Shared briefResearchRiskUserSynthesis
Use when
The task benefits from research, security, user, or commercial perspectives in parallel.
Avoid when
Agents duplicate the same investigation or vote without domain-weighted evidence.
Handoff
Shared brief → isolated findings → explicit disagreements → synthesis decision.
04Shared state / blackboardEvidence ledger + hypotheses + ownership + state transitions
EvidenceHypothesesOwnersStateDecision
Use when
Understanding evolves and several specialists must build on the same observable state.
Avoid when
Every agent can mutate everything, provenance disappears, or stale state looks current.
Handoff
Append-only observations with owner, timestamp, confidence, and supersession links.
05Builder–critic loopDraft → adversarial critique → bounded revision → verifier
DraftCritiqueReviseCloseoutReceipt
Use when
Quality improves through iteration and the acceptance threshold can be tested.
Avoid when
There is no maximum loop count, stop condition, or independent release decision.
Handoff
Artifact hash → findings → exact revisions → closeout verdict.

Working example · cross-house release

One mission. Seven evidence states.

This illustrative workflow uses real public Starlight identities. It shows composition, not a claim that these agents executed a live run.

  1. 01

    Frame

    Human held

    OutputMission brief + acceptance boundary

    GateThe human owner confirms outcome, scope, and irreversible actions.

  2. 02

    Discover

    Parallel

    OutputSource ledger + assumption map

    GateClaims stay source-linked until verified.

  3. 03

    Design

    Synthesize

    OutputSystem contract + experience spec

    GateInterfaces, data, authority, and fallback paths are explicit.

  4. 04

    Build

    Bounded writes

    OutputWorking artifact + change receipt

    GateOne writer owns each surface; external side effects remain held.

  5. 05

    Challenge

    Adversarial

    OutputFindings + bounded revision list

    GateCounterexamples and failure modes are resolved or named.

  6. 06

    Trust

    Fail closed

    OutputSecurity, privacy, and decision verdict

    GateMissing evidence is not a pass; credentials and private memory never cross public handoffs.

  7. 07

    Release & learn

    Receipt bound

    OutputRelease receipt + outcome signal + evolution note

    GatePromotion, public send, spend, and rollback remain owned and observable.

Starlight constellation map

Ten houses. Fifty inspectable profiles.

Adopt a whole house, begin with one conductor and two specialists, or use the profiles as reference contracts for your own system.

01

Sovereign Command

Turn founder intent into bounded, evidence-backed missions while preserving human sovereignty over irreversible decisions.

5 agents
02

Product Forge

Move verified customer problems through system design, bounded implementation, and proof-based quality review.

5 agents
03

Intelligence & Research

Produce source-grounded intelligence, traceable synthesis, and uncertainty-aware forecasts for bounded decisions.

5 agents
04

Creator Worlds

Develop original creator-led stories, visual worlds, audience resonance, and release packages with consent and rights intact.

5 agents
05

Revenue & Venture

Develop evidence-based offers, demand systems, unit economics, partnerships, and venture choices without executing money or binding commitments.

5 agents
06

Community & Academy

Design inclusive learning paths, cohorts, community rituals, facilitation, and evaluation with consent and safeguarding.

5 agents
07

Trust & Safety

Identify, explain, and route safety, privacy, fairness, governance, and incident risks without covert enforcement or offensive action.

5 agents
08

Health & Human Flourishing

Support general wellbeing education, sustainable habits, reflection, movement, nutrition literacy, and care navigation without practicing medicine.

5 agents
09

Enterprise Transformation

Design evidence-backed operating-model, governance, change, AI adoption, and value-realization options without executing organizational authority.

5 agents
10

Web/Spatial/Protocol Futures

Explore accessible web, spatial interface, interoperability, and protocol futures as evidence-labeled prototypes with safe fallbacks.

5 agents

Adoption path

From one workflow to a standing system.

Every step produces an artifact and a proof obligation. Installation comes after composition—not before it.

  1. 01

    Discover

    One recurring workflow and one accountable human owner.

    The current workflow, pain, risk, and baseline are written down.
  2. 02

    Profile

    The minimum agents, skills, non-capabilities, and stop conditions.

    Every role has a unique responsibility and no hidden authority.
  3. 03

    Compose

    A topology, routing map, handoffs, memory scopes, and verifier.

    A failure or unavailable agent has a fallback or graceful degraded path.
  4. 04

    Package

    Digest-bound agent, skill, team, plugin, or marketplace pack.

    The install adapter cannot grant more authority than the portable source declares.
  5. 05

    Pilot

    A bounded shadow or low-risk run with receipts and a stop rule.

    Quality, cost, latency, safety, and human effort are measured—not assumed.
  6. 06

    Evolve

    An evidence-backed version release, migration path, and rollback.

    The new version beats the prior one on declared tests without losing provenance.

Complete experience contract

What an agent system must let people see.

A beautiful profile is only the identity layer. A complete experience makes state, work, trust, memory, evidence, installation, and evolution equally inspectable.

  1. 01Identity

    Name, role, version, status, purpose, visual identity, and canonical source.

  2. 02Mission

    Outcome, owner, constraints, budget, deadline, acceptance, and current stage.

  3. 03Topology

    Active agents, directed handoffs, shared state, dependencies, and fallbacks.

  4. 04Work

    Queue, ownership, progress, blockers, artifacts, and the next bounded action.

  5. 05Trust

    Permissions, tools, data classes, human gates, escalation, and revocation.

  6. 06Memory

    Scope, source, freshness, retention, privacy, contradictions, and deletion behavior.

  7. 07Evidence

    Traces, receipts, evals, costs, decisions, failures, and independent verdicts.

  8. 08Install

    Pack type, digest, dependencies, compatibility, runtime state, and rollback.

  9. 09Evolution

    Lineage, version delta, migration, deprecation, comparison, and learning notes.

  10. 10Access

    Semantic graphs, keyboard operation, text alternatives, reduced motion, and no color-only state.

Five Sovereign Command agents gathered around a luminous mission artifact

Visual expansion · 50-scene production matrix

Give every image a job.

The next visual generation wave is defined before production: five purposeful scene families for each of the ten houses. Exact text, graphs, and UI stay in code.

  1. 01

    Team tableau

    16:9 · .ai house story

    Show the five-agent house as one coordinated system with distinct ownership.

  2. 02

    Workflow in action

    16:9 · .ai workflow explanation

    Make a handoff, tool boundary, or evidence path visible without fake interface text.

  3. 03

    Human gate

    4:5 · .academy trust lesson

    Show the sovereign operator deciding at a meaningful boundary.

  4. 04

    Proof artifact

    1:1 · .org protocol evidence

    Give a pack, receipt, evaluation, or memory object a tactile visual identity.

  5. 05

    Evolution poster

    9:16 · story, social, and release

    Show a version lineage as capability growth with memory and boundaries intact.

Fifty planned visual assets: five scene types for every canonical Starlight swarm.
RunHouseSceneRatioState
01/50Sovereign CommandTeam tableau16:9generated-inspected
02/50Sovereign CommandWorkflow in action16:9planned
03/50Sovereign CommandHuman gate4:5planned
04/50Sovereign CommandProof artifact1:1planned
05/50Sovereign CommandEvolution poster9:16planned
06/50Product ForgeTeam tableau16:9planned
07/50Product ForgeWorkflow in action16:9generated-inspected
08/50Product ForgeHuman gate4:5planned
09/50Product ForgeProof artifact1:1planned
10/50Product ForgeEvolution poster9:16planned
11/50Intelligence & ResearchTeam tableau16:9planned
12/50Intelligence & ResearchWorkflow in action16:9generated-inspected
13/50Intelligence & ResearchHuman gate4:5planned
14/50Intelligence & ResearchProof artifact1:1planned
15/50Intelligence & ResearchEvolution poster9:16planned
16/50Creator WorldsTeam tableau16:9planned
17/50Creator WorldsWorkflow in action16:9planned
18/50Creator WorldsHuman gate4:5planned
19/50Creator WorldsProof artifact1:1planned
20/50Creator WorldsEvolution poster9:16planned
21/50Revenue & VentureTeam tableau16:9planned
22/50Revenue & VentureWorkflow in action16:9planned
23/50Revenue & VentureHuman gate4:5planned
24/50Revenue & VentureProof artifact1:1planned
25/50Revenue & VentureEvolution poster9:16planned
26/50Community & AcademyTeam tableau16:9planned
27/50Community & AcademyWorkflow in action16:9planned
28/50Community & AcademyHuman gate4:5planned
29/50Community & AcademyProof artifact1:1planned
30/50Community & AcademyEvolution poster9:16planned
31/50Trust & SafetyTeam tableau16:9planned
32/50Trust & SafetyWorkflow in action16:9planned
33/50Trust & SafetyHuman gate4:5generated-inspected
34/50Trust & SafetyProof artifact1:1planned
35/50Trust & SafetyEvolution poster9:16planned
36/50Health & Human FlourishingTeam tableau16:9planned
37/50Health & Human FlourishingWorkflow in action16:9planned
38/50Health & Human FlourishingHuman gate4:5planned
39/50Health & Human FlourishingProof artifact1:1planned
40/50Health & Human FlourishingEvolution poster9:16planned
41/50Enterprise TransformationTeam tableau16:9planned
42/50Enterprise TransformationWorkflow in action16:9planned
43/50Enterprise TransformationHuman gate4:5planned
44/50Enterprise TransformationProof artifact1:1planned
45/50Enterprise TransformationEvolution poster9:16planned
46/50Web/Spatial/Protocol FuturesTeam tableau16:9planned
47/50Web/Spatial/Protocol FuturesWorkflow in action16:9planned
48/50Web/Spatial/Protocol FuturesHuman gate4:5planned
49/50Web/Spatial/Protocol FuturesProof artifact1:1planned
50/50Web/Spatial/Protocol FuturesEvolution poster9:16planned

One canon · three public jobs

Build here. Verify there. Learn by operating.

These contracts and examples help people design and inspect multi-agent systems. They do not grant runtime authority or prove live model performance.

Live model evaluation: not_run · Runtime authority granted: false