AWS’s Hydro team wants the compiler to catch coding agents’ mistakes

AWS's Hydro team aims to have a compiler catch errors made by agents in distributed systems. This new approach is described as equivalent to a 5000 line pull request. The team developed Hydro, an open-source Rust framework for distributed systems, where most code is written by agents, with humans making the final decisions. The PRs grew in size and complexity, making it hard to review for the developers who rely on the tool to catch bugs.

Shadaj Laddad, the research lead for the Hydro team, highlighted that code review cannot keep pace with the speed and duration of code written by agents. While tests can't solely rely on catching bugs, Laddad pointed out that the unpredictability of agents makes it difficult for humans to review all the tests and ensure they align with the team's intentions.

To address this, Laddad proposed shared correctness checks outside of the agent's "zone of trust". For example, Rust's borrow checker catches race conditions before the code runs. Hydro employs a similar approach for distributed systems, offering instant feedback to agents about mistakes they made. The team is exploring further domain-specific guardrails that agents can share, starting with patterns identified in historical projects.

Laddad sees this as a new step in a familiar ladder of abstraction. While compilers have limitations, as demonstrated by AI-generated Rust that compiles perfectly but is still wrong, AWS's Infinity agent runtime, launched on October 23, suspends agents waiting for tool calls or CI jobs, enabling 75,000 concurrent agents on a single Raspberry Pi.

Hydro, an early release built for high-throughput, low-latency services, is open source at hydro.run, developed jointly by AWS and UC Berkeley.