AWS Machine Learning vs LiveKit Agents
Side-by-side trajectory, velocity, and editorial themes.
AWS's ML blog has become an agent-pattern catalog built almost entirely on Bedrock.
This feed is AWS Machine Learning blog content, not a product changelog, and it reads as a steady stream of agentic-AI reference architectures. Nearly every recent post composes the same stack — Strands Agents, Bedrock, Bedrock Data Automation, AgentCore Runtime, and MCP servers — into a customer story or how-to. The one genuine release in the window is Agent-EvalKit, an open-source agent evaluation toolkit.
AWS is using the blog to standardize a house pattern for building agents on its own primitives, with document processing and meeting/BI assistants as the recurring demos. Tooling for the unglamorous parts — evaluation via Agent-EvalKit and kernel optimization via Neuron Agentic Development — is starting to appear alongside the showcases. The direction is toward making Bedrock the default substrate teams reach for when wiring agents to enterprise systems.
Expect more of the same composition — Bedrock plus Strands Agents plus MCP — packaged as repeatable blueprints, with additional open-source evaluation and ops tooling to fill the gaps the customer stories expose.
LiveKit Agents 1.6 adds async tools so voice agents stop going silent on long calls.
LiveKit Agents ships at a rapid, release-candidate-heavy cadence, and 1.6.0 lands the headline feature: asynchronous tools that hand control back to the LLM mid-execution and stream progress updates. Between releases the work is steady provider and reliability plumbing across STT, TTS, and the realtime stack, with the usual flow of bug fixes and dependency updates.
The framework is maturing toward production voice agents that stay conversational under real-world latency. Async and cancellable tools, broader STT/TTS provider coverage, realtime model support, and interrupt and turn-handling fixes all point at smoothing the rough edges of live voice interaction. Expect more reliability and provider-breadth work to follow the 1.6 line.
Next releases likely build on the async-tool model with more cancellation and duplicate-call handling, alongside continued STT/TTS provider and realtime-model additions seen throughout these entries.
See more alternatives to AWS Machine Learning →
See more alternatives to LiveKit Agents →