Braintrust JS SDK v3.5.0: Agent Tool Tracing and Experiment Parameters

Here at industry sources, we're tracking Braintrust's continued investment in observability infrastructure for production AI systems. The v3.5.0 release introduces two significant capabilities: Agent tool call tracing for Vercel AI SDK v5 and v6 auto-instrumentation, plus the ability to attach saved parameters directly to experiments. These aren't flashy features - they're foundational for teams running production agents at scale.

Agent tool call tracing extends Braintrust's existing instrumentation to capture the full lifecycle of tool invocations within agentic systems. When your agent calls external tools, each invocation is now automatically traced and logged. This means you can replay exactly what tools were called, in what order, with what inputs and outputs - critical for debugging agent failures in production.

The saved parameters feature addresses a common friction point in experiment iteration. Rather than manually reconstructing parameter sets across experiment runs, you can now attach previously saved configurations directly to new experiments. This streamlines the workflow for systematic parameter tuning without context switching.

• Auto-instrumentation support for AI SDK v5 and v6 reduces manual tracing setup
• Agent tool calls captured with full context for replay and debugging
• Parameter persistence across experiments cuts iteration overhead
• Maintains Braintrust's focus on observability over abstraction

If you're deploying agents built with Vercel's AI SDK, this release directly reduces your debugging surface area. Previously, tracing tool calls required custom instrumentation code. Now it's automatic - you get visibility into agent behavior without writing additional logging infrastructure. For teams running multi-tool agents, this is material time savings.

The parameter attachment feature targets a specific pain point in the experiment workflow. When you're optimizing agent prompts, tool selection, or response strategies, you typically cycle through multiple parameter combinations. Being able to snapshot and reuse successful parameter sets means less manual configuration and fewer opportunities for variation to introduce uncontrolled variables in your experiments.

The underlying message from Braintrust remains consistent: observability and experimentation are inseparable. These updates reinforce that philosophy - better tracing enables better experiments, and better parameter management enables more scientific iteration.

• Reduces custom instrumentation code for tool call visibility
• Enables systematic parameter optimization without manual configuration
• Improves reproducibility across experiment iterations
• Positions Braintrust as the observability layer for AI SDK users

This release reflects growing maturity in the agentic AI development stack. As agents move from prototypes to production systems, the demand for detailed execution tracing intensifies. Tool call tracing is table stakes for production agent monitoring - Braintrust is ensuring they're not left behind as competitors add similar capabilities.

The deeper signal is Braintrust's commitment to working within popular developer frameworks rather than replacing them. By tightly integrating with Vercel's AI SDK rather than building competing abstractions, Braintrust positions itself as the observability and experimentation layer that sits on top of existing tooling. This is the right strategic move for adoption - developers don't want another framework, they want better visibility into the ones they're already using.

Parameter management and experiment tracking are increasingly competitive features across AI observability platforms. Braintrust's ability to attach saved parameters to experiments suggests they're thinking deeper about the workflow - not just logging what happened, but enabling faster iteration. This feature parity will matter for teams evaluating platforms. The momentum in this space continues to accelerate.

• Tool call tracing becoming baseline requirement for production agents
• Integration over abstraction is the winning strategy in agentic AI tooling
• Experiment parameter management emerging as differentiator among observability platforms