pulsehive-runtime

Runtime execution engine for the PulseHive multi-agent SDK.

This crate provides the HiveMind orchestrator, agentic loop (Perceive→Think→Act→Record), workflow agents (Sequential, Parallel, Loop), and the intelligence layer (relationship detection, insight synthesis, context optimization).

Most users should use the pulsehive meta-crate which re-exports this crate alongside pulsehive-core.

Key Components

HiveMind Orchestrator

let hive = HiveMind::builder()
    .substrate_path("my_project.db")
    .llm_provider("openai", provider)
    .embedding_provider(my_embeddings)  // optional
    .build()?;

let stream = hive.deploy(agents, tasks).await?;

Workflow Agents

Kind	Behavior
`AgentKind::Llm`	Single LLM agent with tools and lens-based perception
`AgentKind::Sequential`	Children execute in order, each perceiving previous results
`AgentKind::Parallel`	Children execute concurrently, sharing substrate in real-time
`AgentKind::Loop`	Repeats agent until `[LOOP_DONE]` or max iterations

Intelligence Layer

Component	Purpose
`RelationshipDetector`	Infers semantic relations between experiences
`InsightSynthesizer`	LLM-powered cluster synthesis into derived insights
`ContextOptimizer`	72-hour exponential decay + reinforcement boost
`AttractorDynamics`	Field-level perception warping by high-importance experiences

Perception Pipeline

Each agent perceives the substrate through its Lens:

Query — semantic search or recent experiences
Re-rank — domain weight, type weight, temporal decay, attractor influence
Budget — pack within token limit
Format — render as intrinsic knowledge for the LLM

Performance

Operation	1K experiences	10K experiences
`search_similar(k=20)`	200 µs	279 µs
`get_recent(k=20)`	95 µs	588 µs
`store_experience`	7.0 ms	—
Perception rerank (100 exp)	26 µs	—

License

AGPL-3.0-only

pulsehive-runtime 1.0.1