WORKSTREAM 01Prototype implemented
Data-transfer runtime
A sender and receiver architecture uses multiprocessing workers and zero-copy file transfer over TCP.
RESEARCH IN PROGRESSNO EFFECTIVENESS CLAIMS
Agentic Data Transfer Optimizer
A bounded agentic system that adapts network-transfer concurrency from live TCP telemetry while making every recommendation, constraint, and state transition inspectable.
WORKSTREAM 02Evaluation in progress
Bounded agentic control
Diagnostic and Decision agents reason over live telemetry within hard concurrency limits and oscillation controls.
RESEARCH QUESTIONS
Can an agent change a live transfer without losing control of the system?
The evaluation separates architectural completeness from any claim that adaptation improves performance.
Can agent decisions remain bounded under changing telemetry?
Observe congestion window, retransmissions, RTT variance, and system load while limiting how far concurrency may change at each decision.
When does adaptive concurrency outperform a static setting?
Compare controlled runs only after workload, network path, model, action budget, and measurement provenance are fixed.
SYSTEM PATH
Telemetry enters; bounded concurrency decisions leave.
The architecture exposes where measurement, reasoning, and execution meet.
Transfer runtime
TCP sender and receiver processes move data through a multiprocessing worker pool.
Telemetry
Socket and host observations are captured at each probing interval.
Diagnosis
A diagnostic agent turns current observations and recent state into a structured assessment.
Decision
A decision agent recommends a bounded concurrency change or holds the current setting.
EXECUTION SAFETY
The model recommends. The system enforces.
Hard limits
Concurrency remains inside an explicit permitted range.
Bounded changes
Per-decision deltas and graduated ramping limit abrupt action.
Oscillation control
Repeated reversals are detected and damped before the next action.
What will turn progress into a case study.
Controlled-run provenance, baseline comparisons, repeatable workloads, and public-safe result records must be reviewed before this page makes a performance or optimizer-effectiveness claim.