Agent execution
Dataflow, scheduling, coordination, backpressure, recovery, and continuous execution for multi-step agents.
PROGRAMMING + RUNTIMEResearch Initiative for Delegated Execution / 01
RIDELab
Build the systems agents actually need.
We rethink execution, state, context, evaluation, and operations for long-running agent workloads—not just isolated model calls.
- 04
- research axes
- 01
- shared product
- ∞
- agent horizons
Research agenda
From model serving to agent execution.
Inference engines optimize calls. Agent systems must also manage evolving context, durable state, multi-step control, evidence, and recovery across time.
Context & state
Treat context as a living resource: selected, shared, streamed, versioned, and governed across workflows.
MEMORY + RETRIEVALEvidence & evaluation
Measure behavior, provenance, quality, latency, cost, and safety under realistic repeated workloads.
BENCHMARK + OBSERVEHuman–agent systems
Build controllable, inspectable services that remain useful through changing knowledge, roles, and goals.
INTERACTION + OPERATIONSShared flagship product
SAGE
Streaming-Augmented
Generative Execution
Streaming-computing thinking for LLM inference and agent execution.
SAGE is the shared product and technical vision of the IntelliStream ecosystem. RIDE Lab is the principal engineering steward of its core repositories—not its exclusive owner.
01Explicit dataflowStructure long-running work
02Stateful executionEvolve context over time
03Open boundariesConnect data and inference systems
System portfolio
Ideas become durable systems.
RIDE organizes caller-side agent systems and SAGE product surfaces. Inference internals stay in vLLM-HUST; generic data systems stay in DataSys.
01 / CORESAGE
View every RIDE Lab repository↗
Dataflow-native framework and shared execution core.
FRAMEWORK↗ 02 / AGENTsage-agenticPlanning, tools, workflows, and coordination.
AGENT SYSTEM↗ 03 / CONTEXTsage-ragComposable retrieval and context pipelines.
CONTEXT↗ 04 / EVIDENCEsage-evalMetrics, profilers, judges, and evaluation loops.
EVALUATION↗ 05 / BUILDsage-studioVisual workflow construction and experimentation.
PRODUCT↗ 06 / APPLYSage MateA cited faculty-twin application built with SAGE.
APPLICATION↗Research method
Question. System. Evidence. Transfer.
A project is not finished at the prototype. We stabilize abstractions, publish evidence, and place mechanisms under the organization that can own them long term.
Name the workload and missing abstraction.
Implement against real interfaces.
Measure quality and performance.
Graduate to the correct technical home.
Collaborative ecosystem
One product. Clear ownership.
SAGE connects the ecosystem as a shared product. Each organization keeps a distinct research and implementation boundary.