Agents Need a Save Button: Kitaru's Replay Capability

In the rapidly evolving world of AI agents, a critical gap exists: the absence of a persistent 'save button' that allows developers to revisit and modify past execution states. This limitation hinders the ability to ask crucial 'what if' questions about agent behavior, such as whether a different model or tool configuration would have yielded a cheaper or faster outcome. Hamza Tahir, co-founder of ZenML and builder of Kitaru, highlighted this challenge in a recent presentation, emphasizing the need for a durable runtime layer that can augment agent traces with true checkpoints.

Companies working on this

StartupHub profiles of the companies this article names, with funding and a one-liner from our database.

DoorDash

$16.0B

DoorDash is the largest food delivery platform in the US, connecting consumers with restaurants and convenience stores through on-demand delivery.

DoorDash

$16.0B

DoorDash is the largest food delivery platform in the US, connecting consumers with restaurants and convenience stores through on-demand delivery.

Ole

$2M

Fashion e-commerce platform delivering clothes from top stores within 50 minutes.

Tahir explained that while traces offer valuable telemetry data on tool calls and their inputs/outputs, they are inherently disconnected from the agent's actual runtime environment. This means that variables, in-flight file systems, decision-making logic, and the code itself are lost once the trace is emitted. This disconnection prevents developers from precisely understanding why an agent behaved a certain way or exploring alternative execution paths without rerunning the entire process from scratch.

The Power of Replay: Unlocking 'What If' Scenarios

The core of Tahir's argument centers on the concept of replayability. Just as users have been able to save and revisit their work in documents since the 1980s, agents need a similar mechanism. A 'save button' for agents would enable them to store their state at various checkpoints, allowing for the re-execution of specific segments with modified parameters. This capability opens up a range of powerful analytical and debugging scenarios:

• Model Swapping: Test the impact of using cheaper, open-source models by replaying an execution with a different model selected at a specific checkpoint.
• Tool Mocking: Override the output of a tool to simulate different conditions or isolate specific functionalities.
• Degradation Testing: Intentionally degrade a tool's performance to observe how the agent handles errors or unexpected behavior.

Tahir introduced Kitaru, a new tool from ZenML, as a solution that bridges this gap. Kitaru provides a durable runtime layer beneath existing agent frameworks, augmenting traces with code execution details and environmental context. This allows for the creation of persistent checkpoints that capture the complete state of an agent's run, making replay and 'what if' analysis feasible.

From Production Traces to Actionable Insights

The true power of this approach lies in its ability to leverage production data. Tahir argued that production runs serve as the ultimate test set, capturing real-world inputs, edge cases, and actual tool responses. By having a system like Kitaru in place, organizations can mine their existing production traces to perform detailed analysis and drive improvements. For instance, an agent handling customer refunds after chargeback disputes could be analyzed to see if language variations or escalation decisions could have been handled differently, potentially leading to cost savings or faster resolution times.

The process, as Tahir outlined, involves a four-step methodology: Checkpoint, Replay, Diff, and Decide. By checkpointing key runs (expensive, failed, or risky ones), replaying them with a single change, diffing the results, and then making informed decisions about shipping, routing, or holding, teams can systematically optimize their agents. He cited DoorDash's success with a similar simulation and evaluation platform for their support chatbots, which reportedly reduced hallucination rates by 90% and cut down testing time from hours to minutes.

The Importance of Cohort Analysis and Avoiding False Economies

Tahir cautioned against relying on single replays, emphasizing that "one replay is an anecdote, ten is evidence." He highlighted that models passing tasks 60% of the time might only be self-consistent a quarter of the time, underscoring the need for cohort analysis to understand the broader distribution of outcomes. A naive model swap, he warned, can lead to a 'false economy' if it appears cheaper on paper but degrades the overall value created by the agent.

The ultimate goal is to close the loop on agent development and evaluation, turning past production runs into a continuous cycle of improvement. By integrating replay capabilities into the release gate process, teams can ensure their agents are not only cost-effective but also robust and reliable, ultimately leading to better, cheaper, and faster systems.

The presentation concluded with a call to action: "Wrap your agent. Start replaying." Kitaru is available as an open-source, self-hosted solution, providing the necessary infrastructure for production ML agents to achieve this crucial level of introspection and optimization.