AIFA Daemon Loom
Managed orchestration plane: intake, policy, gates, MCP/API surfaces.
D-Loom · textile runtime
Daeloom
execution without session bloat
D-Loom is the token-efficient discipline inside AIFA Daemon Loom: many small, disposable Needles coordinated by a Main Thread, with state in the Spindle and tasks in the Pattern Ledger — not in every prompt.
What is D-Loom?
The broader Loom is an orchestration fabric for durable agent workloads. D-Loom is how that fabric runs work efficiently: do not keep agents alive longer than the process they were created to complete. That means less context injection, fewer idle sessions, and auditable progress through explicit task state instead of chat history.
On ClusterClaw, Daeloom is the execution story for the enterprise AIFA add-on — same hosted OpenClaw platform, extended with always-on orchestration when you need multi-agent operations at scale.
How a run moves
- 1
Plan on the Main Thread
One lean orchestration session holds the Pattern Ledger, picks the next task, and dispatches work — it does not do every micro-step itself.
- 2
Spawn Disposable Needles
Each atomic unit of work gets a small agent with a Context Diet: only what that task needs, then a structured check-in.
- 3
Process-completion TTL
Agents are not timed out by the clock. They exit when their process hits a terminal state: completed, failed, blocked, escalated, or cancelled.
- 4
Update the Pattern Ledger
Human- and machine-readable task state lives in the ledger. Needles annotate only their row; the Main Thread decides what runs next.
- 5
Join, then continue
Parallel Needles report to a join barrier; results merge into the Spindle before the next fan-out.
Plan → Spawn Needle → Execute → Check-in → Update Ledger → Exit → Continue
Textile vocabulary
Metaphors map to real runtime concepts — readable for humans, precise for agents.
- Thread
- One persistent workload lifecycle (a run you can resume and audit).
- Needle
- A scoped execution agent that carries a Thread through one action.
- Disposable Needle
- Short-lived agent: minimal context, one task, check-in, exit.
- Spindle
- Continuity layer — checkpoints and summaries, not bloated prompts.
- Reed
- Concurrency control — how many Needles run in parallel.
- Thimble
- Safety layer — policy, approvals, and guardrails.
- Weave
- The composed fabric of threads, signals, and delivery.
Design laws (plain language)
- TTL is process-completion-based, not a timer on the wall.
- Prefer many tiny Needles over one agent carrying the whole repo.
- Task truth lives in the Pattern Ledger, not in chat memory.
- Summaries and checkpoints live in the Spindle, not every session.
- Parallel work uses fan-out and a join barrier before the next step.
- Needles exit after terminal check-in — no orphaned sessions.
D-Loom runtime
Disposable Needles, Main Thread, Context Diet, Reed-controlled concurrency.
ClusterClaw
Hosted OpenClaw gateways and isolation lanes — Cloud Run default, GKE enterprise.
Need a hosted operations plane?
Daemon Loom Cloud is where developers drop off durable workflows — API, MCP, console, webhooks, and audit history. Same D-Loom discipline, surfaced as a product on ClusterClaw lanes.
Daemon Loom Cloud on daeloom.cloudRoadmap
Architecture and loom-state contracts are in-repo; hosted D-Loom orchestration is not wired to production yet. We are moving from validation and sidecar prototypes toward API/MCP alpha and cloud beta with enterprise controls.