Meterless Swarms v0.1.9 - End AI Sameness

Plan. Parallelize. Merge. Verify. Ship stronger work.

Swarms is a local-first, browser-native multi-agent execution engine. Give it one task and it turns that task into a coordinated system that plans, distributes work, runs specialist agents in parallel, merges outputs, verifies quality, and returns a stronger final result than any single agent could produce alone.

The shift

Prompt → Model → Subagents → Simple Handoffs →  Output           ← old
Prompt → DAG Plan → Parallel Agents → Merge → Verify → Output   ← Swarms

What's new in this release

• Hardened DAG integrity — every plan and dynamic agent spawn is validated up front: cycle detection (Kahn's), dangling depends_on / edge-endpoint rejection, self-loop prevention, and full schema validation. A task that depends on something that doesn't exist is rejected before it can wedge the scheduler.
• LLM multi-stage DAG planner — an opt-in ARCHITECT pass decomposes a prompt into a real dependency graph with synthesized agent specs; the proven flat fan-out remains the always-available fallback.
• Topological scheduling — named Kahn ordering with parallelism-level analysis exposes the true peak concurrency width of a run.
• Provider failure classification + retry — transient errors (HTTP 429 / rate-limit / timeout / 5xx) are classified (rate_limit / quota / transient / fatal) and retried with exponential backoff and full jitter, so a single throttle no longer silently kills a variant. Quota/fatal errors fail fast instead of masquerading as task failures.

What's inside

• DAG planning — prompts become validated task graphs (cycle, reference, and schema checks) before anything runs
• Parallel agents — up to 64 specialists concurrently, with per-task model assignment, provider throttling, and classified retry/backoff
• Blackboard coordination — agents share intermediate outputs instead of working blind
• Merge phase — swarmMerger reconciles artifacts and resolves conflicts into synthesis
• Verification phase — swarmVerifier QA agents validate completeness before output ships
• Markovian engine — @@@STATE@@@ carryover compresses state across chunks so long tasks scale
• H-MEM 4.0 — short-term, working, and long-term memory across sessions and runs
• Local-first runtime — runs entirely in your browser; an IndexedDB (V7) store persists runs, artifacts, telemetry, files, and memory. Inference routes to your configured cloud providers (a WebLLM / WebGPU local path is scaffolded in the registry but not yet enabled at dispatch)
• Live artifact refinement — surgically edit selected HTML previews and Markdown outputs instead of regenerating everything
• Replay + trust ledger — runs and memory operations are inspectable

Core Functionality

Swarms is one focused workflow the spins up varient approaches to any goal. You type a single prompt and the app runs the full plan → parallelize → merge → verify pipeline, surfaced through:

• Grid view — every variant/artifact a run produces, as selectable tiles
• Live DAG view — the task graph animating per-task state (queued / running / done / failed) in real time as the scheduler works
• Artifact refinement — click-to-select editing of HTML previews plus a Markdown refine surface, so you can revise one output instead of regenerating it
• H-MEM panel — browse memory alongside Dream, Conflicts, Sleep, and Trust Ledger workflows
• Markovian engine — @@@STATE@@@ carryover for long, chunked tasks
• Scheduled runs — queue and replay swarm runs on a recurrence
• Settings — add provider API keys and pick the model for each routing slot (Code-Gen / Architect / Research / Mission)

Glass-box orchestration

The task DAG, agent assignments, prompt inputs, model choices, intermediate artifacts, telemetry, run history, merge decisions, and verification reports are all inspectable. Visible workforce, not hidden black box.

Architecture

React 19 + TypeScript + Vite, browser-native client-side runtime. Persistence: IndexedDB (V7 schema) + LocalStorage compatibility. Providers: Google, OpenAI, Anthropic, OpenRouter, OpenCode Go, and any custom OpenAI-compatible endpoint (Ollama, LM Studio, vLLM, …); a local WebLLM path exists in the registry but is not yet enabled at dispatch. Orchestration subsystems: swarmPlanner, swarmScheduler, swarmAgentRunner, swarmMerger, swarmVerifier, swarmSpawnDecisionEngine, swarmGraphUtils, swarmSchemas, providerFailureClassifier.

About Meterless

Swarms is a standalone product. Meterless also publishes Gaia and Relay as separate, independent applications — each solves a different problem and can be used on its own. Pick the one that matches what you need; you don't need the others to get value from Swarms.

Run state survives

Runs persist locally and are designed to survive browser crashes, closed tabs, and restarts — work continues or is inspected later.

See the README and LICENSE.md for details.