OrcaRouter announced OrcaReplay, a new open-source record-and-replay engine designed to make AI agents reproducible, debuggable, and easier to understand.
OrcaReplay captures agent execution as a replayable trace. Developers can inspect what the model saw, which tools were called, what changed in the environment, and where the execution diverged. From any recorded step, developers can replay the execution, debug model decisions and tool calls, or fork the run and test a different model, prompt, tool result, or decision path without restarting the entire workflow.
This makes it possible to answer questions such as:
- Why did the agent make this decision?
- Which model or tool call caused the failure?
- Can the failure be reproduced?
- Would another model succeed from the exact same state?
- Where did two agent runs begin to diverge?
OrcaReplay is also designed for the nondeterministic nature of agent systems. Shell commands, external APIs, filesystem changes, and network responses may create side effects that cannot safely or reliably be executed again.
During replay, OrcaReplay can reuse recorded outputs and captured artifacts instead of blindly repeating those actions. Developers can then explicitly choose when a fork should interact with a live environment.
OrcaReplay is fully open source, agent-agnostic, and model-agnostic. Replay data can remain local, making it suitable for privacy-sensitive development, debugging, evaluation, and security workflows.
The project can be used independently or together with OrcaRouter, which provides a unified API for accessing and routing across a broad catalog of frontier and open models.
OrcaReplay is available now:
The Latest
For decades, traditional observability tools have forced IT professionals to sift through endless dashboards, turning a flood of alerts into actionable insight and driving widespread burnout. While enterprises adopted AI to eliminate this friction, most still deploy it reactively rather than proactively ... As systems become more complex and IT moves to the center of business strategy, a reactive posture is unsustainable ...
Ask an operations team what they monitor and you will hear a familiar list: latency, errors, throughput, uptime. Those signals still matter for AI, but they no longer answer the most important question. As Krishna Sai, CTO of SolarWinds, puts it: "Traditional observability tells you a system is running. AI observability tells you whether it's still reasoning soundly. Those are completely different questions" ...
AI observability serves several roles across an organization. AI engineers use it to build, evaluate, and improve AI systems. SREs use it to operate those systems reliably and efficiently in production, including managing their cost. Leaders need it to connect AI behavior and operational performance to business outcomes. The common need is a clear understanding of whether AI systems are delivering the intended result ...
Part 4 outlines the challenges enterprise face when adopting AI observability. The main challenge is that AI complexity compounds at every stage of maturity. LLM-based applications introduce new behavior, dependencies, and quality considerations, while Agentic AI adds further failure modes as systems make decisions and take actions ...
Part 3 reports on the current state of AI observability. Most experts agree: enterprises are moving faster with AI than they are with the operational discipline needed to run it safely ...
When discussing AI Observability, one of the first questions that arises: Is this just an extension of traditional observability or is this a whole new type of tool? Part 2 compares traditional observability with AI observability ...
In this extensive 14-part series that will post over the next three weeks, experts across observability and AI will provide insight and opinions on the need for AI Observability, the challenges, the benefits, the current state, and predictions for the future ...
Amid rising complexity, organizations are moving quickly toward agent-driven operations: AgenticOps, according to The Impact of Agentic AI on Network Operations, a new report from Cisco ...
AI, automation, and enterprise software give organizations an extraordinary amount of leverage. Work that once required hours can happen in minutes. Information buried across systems can be analyzed almost instantly. Complex workflows can be coordinated across teams with far less manual intervention. That leverage creates a temptation to start with the technology ...
Everyone wants to talk about models. Compute clusters. Benchmark scores that climb a fraction of a point and make headlines anyway. It's an easy conversation to have because it's measurable. But it's also increasingly beside the point ...