AWS gives AI agents a way to investigate Lambda execution history

AWS has launched a Lambda telemetry exporter that records durable functions’ step-by-step execution history and sends it to as many as 13 destinations. A companion Model Context Protocol server lets AI agents query that data in natural language, turning operational traces into something an on-call engineer can investigate by asking a question.

AWS AI Watch analysis

What happened

The exporter captures details including execution status, operation timing, step inputs and outputs, and errors. AWS says customers can send the data to destinations including CloudWatch Logs, Amazon S3, DynamoDB, Aurora and OpenSearch, and keep the history beyond the standard retention period.

The companion MCP server can let an agent such as Kiro find recent failed executions and investigate their causes. The exporter supports TypeScript, Python, Java and Go. AWS says customers can redact sensitive fields before export and set a sampling rate to manage data volume. It is available at no additional charge in Regions that support Lambda durable functions. AWS’s announcement has the setup details.

Why it matters

This puts an AI query interface on top of execution history that teams can already route into their observability systems. Instead of manually chasing failures across logs, an engineer could ask an agent to find recent problems and surface likely causes. The underlying telemetry also remains available for longer-term operational or business analysis.

That is a useful, concrete job for an AI agent: helping people navigate operational data, rather than asking it to run the infrastructure unsupervised. The practical value will depend on how well the exporter fits into a team’s existing monitoring and access controls.

Our read

There is more here than an MCP connector bolted on for fashionable reasons. The exporter gives teams control over destinations, retention, redaction and sampling, while the companion server offers a direct route from a question to execution history. AWS says the exporter carries no additional charge, though normal costs for the chosen destinations and their use may still apply. For teams running durable Lambda workflows, this is worth a look if tracing failures is currently a slog.

What to watch

• Which telemetry destinations and agent setups teams use in practice.
• Whether natural-language queries reliably surface useful causes, rather than merely shortening the hunt through logs.
• How teams configure redaction, sampling and access when agents can query execution data.

Discussion spark: Would you let an AI agent investigate production execution history, or should it only point engineers to the relevant logs?

not affiliated with or endorsed by Amazon Web Services (AWS)