- Who it helps
- Inference engineers measuring scheduling, caching or serving performance.
- Availability
- The September 26 paper is public. A working implementation, supported hardware and complete reproduction cost were not verified.
Preserve what affects serving cost
The paper describes recording token sequences, expert routing, tool durations and dependencies. Replay keeps autoregressive computation while forcing the recorded tokens. This is intended to preserve workload details that simple request-length simulations can lose, such as cache behavior or routing through different model experts.
Separate performance replay from app debugging
This evidence concerns inference-serving comparisons. It does not establish a tool that recreates arbitrary application state or reproduces every agent bug. For that broader use, a developer would also need the external tool responses, permissions and state transitions that caused the failure. Those capabilities were not verified here.
What you can do
- Read what the trace captures and check whether those fields match the performance variable you intend to compare. Open reference ↗
Where the evidence stops
The paper abstract was read; replay code and performance claims were not executed or independently reproduced.
Sources read. Sources checked October 22, 2026 UTC. No external software workflow, model inference or reported experiment was executed.
Sources & further reading
Paper · paper
AgentReplay: reproducible agentic serving workloads ↗
Original abstract read. Replay targets serving-system performance with recorded generation and dependencies; application-state replay was not established.
AgentReplay authors · page checked 2026-10-22 · Published 2026-09-26
Coverage note: The original abstract supports the serving-workload explanation. Implementation, measured performance and independent discussion remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
AgentReplay records the generation and dependency structure of an agent workload so serving-system changes can be compared on repeatable work.