post-launch wave-3: multi-agent crew domain (shared Φ + co-selection)

[DubovskiyIM]

Why

Wave-1 (invest) demonstrates a single agent acting on its own preapproval. Wave-2 (deploy-pipeline, see #3) demonstrates one agent + one human reviewer. Wave-3 should demonstrate the next problem layer: multiple agents coordinating in the same domain without stepping on each other.

This is the structural gap that's eating CrewAI / AutoGen / LangGraph use-cases right now (see research findings in ~/Desktop/IDF/2026-05-03-ai-angles-research-spec-runtime-mcp.html): multi-agent frameworks orchestrate agent calls, but they don't manage shared state between agents. The community currently solves this with custom state-stores, vector-databases, or framework-specific scratchpads. None of those carry domain semantics.

Fold already has the primitive: Φ as event log + co-selection cross-projection state-sharing. Co-selection landed in idf-sdk PRs #300/#303/#308/#311/#313 in April 2026. We just haven't shipped a domain that uses it as the demo's core narrative.

Domain candidates

Pick one from below. Lean: incident response — recognizable HN audience, natural multi-agent role split, real-world consequences.

Option A — incident response (recommended)

Three agents + one on-call human in the same Incident lifecycle:

• Triage agent — classifies severity, finds related runbooks
• Investigator agent — pulls logs, traces, correlates with recent deploys (via cross-domain link to wave-2 deploy-pipeline)
• Mitigator agent — proposes rollback / config-change / scale-up
• On-call human — approves mitigation; only human can close_incident

Shared state: Incident.timeline[], Incident.status, Incident.assignedTo. Co-selection lets all three agents see what the others are looking at without polling.

Option B — sales pipeline

CRM-like domain. Agents specialize: enrichment, follow-up, scheduling, contract-drafting. Shared Deal state. Less HN-resonant (audience is dev, not sales), but closest to existing CrewAI demos — direct apples-to-apples comparison signal.

Option C — code review

Three agents: linter, security-scanner, architect-reviewer. Single PR, shared verdict. Closest to GitHub-native HN audience, but the «multiple perspectives on one PR» model is well-trodden — competitors include Greptile, CodeRabbit, qodo.

Three rejection types this domain showcases

1. Co-selection conflict — two agents try to mutate the same Incident.assignedTo simultaneously. The runtime intercepts the second mutation with a structured rejection citing the first agent's lock. Cardinality invariant: max 1 active assignee.
2. Cross-agent SoD — Investigator agent cannot also Mitigate the same incident (separation-of-duties expression invariant: mitigator !== investigator). Mirrors the wave-2 approve_deployment SoD pattern.
3. Dependency-graph constraint — Mitigator cannot propose_rollback until Investigator has logged at least one correlation_finding. Cardinality invariant with where: { type: "correlation" }, min: 1.

What's new vs Wave-1/2

• Co-selection demo — first quickstart that exercises <CoSelectionProvider> + useCoSelection. Not just runtime guards, but shared cursor / focus between agents-as-pixels (when wired to UI). For the npm/MCP demo, manifests as agents reading each other's currentFocus.{incidentId, fieldPath} from /world.
• Cross-domain reference — Incident.deployTriggeredBy → Deployment.id (FK to wave-2 domain). Demonstrates that IDF artifacts compose across domains without runtime coupling.
• First multi-agent preapproval — each agent has its own AgentPreapproval row, with allowedAgentRoles: ["triage"] field. The runtime's preapproval guard now needs to check role-scoped action sets, not just user-scoped.

Implementation paths

Path A — preset preset in existing Docker quickstart (~2 hours):

• New idf repo branch with domains/incident-response/{ontology,intents}.js
• BOOTSTRAP_DOMAIN=incident docker compose up
• Three agent-personas as separate test users (entrypoint.sh seeds them)
• New scripts: demo:co-selection-conflict, demo:sod-cross-agent, demo:dependency-graph

Path B — full multi-agent simulator (~1-2 days):

• Background process in container that simulates Triage / Investigator / Mitigator concurrently with setInterval calls. Real LLM-driven multi-agent demo without requiring viewers to set up CrewAI / LangGraph themselves.
• Demo shows agents competing for state, runtime resolves via Φ's proposed → confirmed | rejected semantics + co-selection.

Lean: A first, B as wave-3.5 if A gets traction.

Open questions

1. Co-selection requires <CoSelectionProvider> mounted in a UI client. For npm/MCP demo, what's the equivalent? Maybe /api/agent/:domain/focus GET endpoint that returns last-N agents' current focus rows. Schema TBD.
2. Should the human in this demo be an oncall role or incident-commander? Tradeoff: oncall is recognizable, incident-commander is more accurate for governance. Lean: oncall for demo-clarity.
3. Cross-domain link to wave-2 — only works if wave-2 is shipped first. Hard dependency: wave-3 cannot demo before wave-2.
4. Is it 3 agents + 1 human, or N agents? Tradeoff: 3 is concrete, N is impressive but harder to demo in 75 sec. Lean: 3.

Decision frame

Open this only if:

• Wave-2 (post-launch wave-2: bundled deployment-pipeline demo domain + npx fold-demo #3) shipped and validated post-launch
• HN-thread or X-thread surfaced explicit «what about multi-agent» asks
• We have the bandwidth to also handle the cross-domain composition demo (which is the real wow-shot of wave-3)

Don't open this:

• If wave-2 metrics are lukewarm — pivot to compliance (fix(docker): убрать image: line — compose не пытается pull несуществующий образ #5) or decision lineage (docs: add Loom walkthrough в README #6) instead
• If LangGraph/CrewAI ship native shared-state primitives in Q3 2026 (then we're commenting on top of their stack, not differentiating)

Source narrative

Wave-3 closes the gap CrewAI / AutoGen / LangGraph leave open: agent orchestration vs agent substrate. Frameworks tell agents what to do next. Fold tells agents what world they share.

«Multi-agent isn't orchestration. It's a shared substrate.»

This issue records the design before it's needed. Decisions made today (3 agents > N, incident-response > sales, Path A > B) are revocable when issue is reopened.