Built in phases

The foundation this site describes: the envelope protocol over 4 topic exchanges, HMAC-signed identity, the retry contract, the orchestrator control plane — and a catalogue of capability agents for documents, NLP, media, security scanning, data pipelines and AI relay, including choreographic workflows like ticket enrichment and the weekly compliance report.

Measured: 14 agents and 11 adaptors registering by capability · 4 exchanges · 42 fabric services

The workflow layer, running today: named, versioned graphs authored visually, fired six ways, executed by a dedicated runtime with per-node state — exercised end to end by the sensing estate: multimodal sensors, enrichment at ingestion, anomaly detection, fusion agents and live scenario consoles in Prism.

Measured: 69 built-in nodes · 30 bundled workflows in 12 families · 30 telemetry tables across 10 source schemas

The operator surface: agent health and queue depths at a glance, the live audit feed with full request traces, the dead-letter viewer with inspect-and-replay, the transducer studio and its run history, drafts awaiting approval. Most of the fabric is invisible; the consoles are where all of it is observable — and actionable.

Measured: 4 bundled consoles · 60 panel data sources · 13 Cortex representations

Every token with an owner: AI calls priced as they happen against operator-editable rules, landed in per-tenant time-series with continuous aggregates, native-currency reporting and per-tenant billing modes — bring your own keys or pay platform rates. Chargeback isn't a quarter-end spreadsheet; it's a query.

Measured: 17 hypertables · 45 continuous aggregates · 10 provider routes resolved per tenant

Loom speaks the open agent protocols — A2A at the fabric edge, MCP for tools — and turns any agent, in any language, into a first-class citizen through a thin adapter. Bring an A2A Agent Card, an OpenAPI service, a Microsoft 365 manifest or a framework export; Loom mints a capability, binds the runtime and the agent inherits identity, retries, audit and cost like any native one. See how it works →

Cortex consumes the audit stream and re-presents the fabric through many lenses at once: rate counters, semantic embeddings, trends and anomaly scores, aggregated over windows and projected into navigable 2D and 3D spaces — activation matrices, embedding clouds and correlation views on the operator dashboard today. It also lays the groundwork the later stigmergic layer will sense: a pheromone field in shared state, already being written.

Describe a goal; the planner composes a candidate transducer from the capability catalogue, validates it structurally and persists a draft for operator approval. The core loop runs today — the Q4 work is production-hardening it: richer capability reasoning, draft review in the studio, and every published graph teaching the next draft.

The experience layer meets the fabric: portal dashboards whose panels are enriched by Loom agents — pre-reads prepared before you open them, watches that follow your scope, suggested offloads — all opt-in, all running as a named principal over the same signed envelopes, so every enrichment is attributable and auditable.

The first console of this kind is running: the Demo Wright, a Loom console that interviews a requester, researches the estate, drafts the demo definition and walks it through mechanical compliance gates and a human sign‑off — then hands the build to the Substrate forge, where a machine is minted with a time‑to‑live and the live demo is verified beat by beat. A conversation becomes a specified, gated, forged and verified product. The factory this feeds →

Elasticity isn't Loom's to deliver — it's Micro‑Substrate's. Cortex already watches queue-depth growth, throughput collapse and timeout rates; this phase feeds those signals down to the substrate, which owns the machines and answers with capacity — more replicas when backlogs grow, fewer when they drain. Intelligence senses, the foundation flexes: the first closed loop between the layers. Advisory first; automatic once trusted.

The swarm layer goes active: agents that coordinate indirectly through the environment — sensing and reinforcing pheromone trails in shared state rather than being commanded — closing the feedback loop down into the substrate itself, so infrastructure signals shape behaviour and behaviour shapes infrastructure. Acting only ever through Phase 1 agents, with recommendations before actions.

The longer bet, and where the phases converge: the goal planner, the pheromone field and the stigmergic layer growing into a colony — a fabric that notices patterns in its own behaviour, proposes new workflows for them, and improves the ones it already runs. Emergence welcome, unsupervised mutation not: operators always hold the publish button.

A fabric that only adds is a fabric nobody prunes. What has been taken off the loom, and what replaced it:

• fabric-colony-planner · 2026-06-02 → fabric-agentic-planner. The pipelined six-stage planner capped catalogue-gap diagnosis at 1 of 19 unsolvable goals.
• five per-task AI agents (sentiment, classifier, embeddings, summariser, translator) · 2026-05-14 → one fabric-agent-ai. Thirteen AI tasks behind one tenant-aware client instead of five fat containers.
• fabric-iot-aggregator → fabric-stream-aggregator. One aggregator for every stream, not one per device class.