The Mezmo team developed "Aura," a Rust-based harness, to solve the challenges of incident response in SaaS operations dealing with petabyte-scale data. They stated that traditional LLM utilization had led to issues such as context overflow and difficulties in permission management.

Aura utilizes a system where workers are defined for specific task domains, such as log reviews and metrics analysis, to conduct collaborative investigations. Permissions and tool access are defined in code, ensuring deterministic control so that agents cannot arbitrarily acquire capabilities via prompts. Additionally, it executes sensitive operations through Human-in-the-Loop (HITL) approval and features fault tolerance.

Significant effort was put into managing the context window; large tool outputs and worker responses are saved to disk. A tool is provided that allows the agent to read only the necessary slices of this data. The project is released under the Apache 2.0 license and can be installed via rpm, deb, and brew.

In a demo using DeepSeek-V2-Flash, Aura identified 502 errors in a checkout service as a memory leak. The process demonstrated the agent pinpointing a defect in a recently merged PR and proposing a fix on GitHub. Throughout this process, an audit trail is maintained via OTEL events.

Technically, an agent coordinator drafts and executes DAG flows, allowing workers to perform parallel investigations. Workers use strongly typed persistent attributes to suppress redundant recalculations of remote data. The system also supports header routing that avoids direct key access and HITL interruptions via Webhooks.

Currently, the asynchronous input system for running as a service is under development. Furthermore, because there is no interruption mechanism via incoming Webhooks, middleware or MCP polling is required for automation. The development team is seeking feedback from users and contributors regarding future directions.

Source: Show HN: Aura – a Rust agent that investigates and fixes production incidents (Hacker News Frontpage, 2026-09-03)