The Spark for AI agents.
Multi-agent systems break at scale. KiloAgents is the runtime that makes parallel AI agent workflows actually work.
Backed by Y Combinator S26
The multi-agent scaling wall
Coordination overhead grows super-linearly
Google ResearchMaximum effective agents before degradation
Industry benchmarksToken explosion in multi-agent systems
Stanford HAITeams cite inter-agent communication as #1 latency source
Developer survey, 2025We've seen this before
Rigid two-stage pipelines
Rigid orchestrator-worker patterns
Disk-based, no shared state
No shared context between agents
No execution optimization
No workflow optimization
Then Spark arrived.
Now KiloAgents.
How it works
Versioned Context Snapshots
Immutable, tracked context objects — like RDDs for agent state. Full lineage tracking so you can trace any decision back to its source.
Task Dependency DAG
Automatic parallelism detection. We build the graph, identify what can run concurrently, and sequence dependent tasks correctly.
Workflow Optimizer
The missing piece. No agent framework has this. We analyze the full task graph and eliminate redundant operations before any agent starts working.
Shared Fast-Access Context Layer
Tiered context (working, session, memory) with compiled views. Each agent gets exactly the relevant slice — 7.8x speedup through intelligent cache sharing.
Lineage-Based Fault Tolerance
When an agent fails, recompute only the affected downstream tasks. Not everything. Just like Spark recomputes lost RDD partitions.
Performance at scale
Benchmark: Multi-agent research + synthesis pipeline, measured on 8-core cloud instance
Token reduction
Latency improvement
Effective agent scaling
Architecture
Quanlai Li
Founder & CEO
I built BIDMach at Berkeley — GPU-accelerated ML that outperformed 50-node Spark clusters. Now I'm applying the same systems thinking to AI agents.
The runtime layer is missing.
We're building it.
Open source core. Managed cloud coming soon.