SwarmDev
A deterministic pipeline that turns a request into an inspectable plan.
Prompt classification, planning, graph construction and execution orchestration, with a diagnostic trail at every step.
Moving Beyond the Single Opaque Model Call
SwarmDev is the conversational coding and research pipeline used inside AutoML App. It takes a natural-language request and works through it in explicit, repeatable stages rather than a single opaque model call.
Each stage produces artifacts you can read, test and diff, which makes the pipeline significantly easier to debug, audit, and benchmark than a black-box agent.
Every stage outputs structured JSON/typed frames ready for diffs.
Consistent inputs yield deterministic execution graphs.
Explicit pause, hold, and human escalation checkpoints.
How It Works
Six explicit, sequential stages decompose ambiguous user intent into verified execution plans and plain-English summaries.
Classify
Intent & Workflow IdentificationIdentify the intent and workflow family of the incoming natural-language request, selecting the appropriate plan schema.
Analyze
Syntactic & Semantic Signal ExtractionSplit the request into discrete syntactic tokens and semantic signals, isolating entities, constraints, and dependencies.
Graph
Knowledge & Planning Subgraph ConstructionBuild an explicit knowledge graph and an executable planning subgraph from the extracted signal frames.
Route
Triple-Frame Dispatch & Conflict ResolutionSend triple frames to specialized execution handlers and resolve structural conflicts or edge ambiguities between them.
Normalize
Token Standardization & Policy GatingStandardize tokens and enforce hard policy gates, validation checks, and safety invariants before execution.
Emit
Artifact Generation & Natural-Language RenderingProduce execution, validation and telemetry artifacts, and render a plain-English response from the final response graph.
Core Capabilities & Substrates
Engineered for high-assurance environments requiring reproducible planning and modular reasoning layers.
Deterministic core
Same input produces the exact same plan, allowing complete end-to-end regression testing and deterministic CI validation without model stochasticity.
Graph reasoning
An optional kGNN layer with adjacency indexing, weighted message passing, and embedding-ranked cutout resolution for deep topological inference.
Swarm-guided node selection
An optional bridge to HyperSwarm's coherent wavescan for choosing optimal graph nodes. Both this and the kGNN layer are opt-in via service registration.
Policy gates and escalation
Requests can be held, checkpointed, or escalated to supervisor review instead of silently executing risky or ambiguous operations.
Diagnostics and sidecar
Built-in structured telemetry stream and a dedicated sidecar interface for real-time observation, profiling, and debugging from external developer tools.
Contracts-first design
Schemas and behavioral contracts are strictly decoupled from implementation details, ensuring all stage components and handlers are modular and swappable.
Runtime Status & Integration Target
Status: Core component of AutoML App, exposed as a modular pipeline library. It runs inside AutoML App as a node integration.
Host Environment
Runs as a managed library inside AutoML App with zero disk overhead and direct in-memory message framing.
Application Node
Embedded as a core execution and research node inside AutoML App for autonomous machine learning experiments.
Opt-in Extensibility
The kGNN graph reasoning engine and HyperSwarm coherent wavescan bridge are activated via standard dependency injection / service registration.
// Register SwarmDev deterministic pipeline in AutoML App
services.AddSwarmDevPipeline(options =>
{
options.EnableDeterministicPlanning = true;
options.EnableDiagnosticsSidecar = true;
// Opt-in Graph Reasoning (kGNN with adjacency indexing & weighted message passing)
options.UseGraphReasoningLayer(kGnn =>
{
kGnn.MaxHops = 3;
kGnn.CutoutResolution = CutoutResolutionMode.EmbeddingRanked;
});
// Opt-in Swarm-Guided Node Selection via HyperSwarm Coherent Wavescan
options.UseHyperSwarmWaveScanBridge();
// Register hard safety & policy gates
options.AddPolicyGate<ArchitecturalInvariantGate>();
});Explore SwarmDev in the HFLabs Ecosystem
Review the full operator manuals, explore Hypertensor GraphRAG, or inspect the polyglot matrix.