AWS wants the compiler to catch what agents get wrong in code

The AWS research team behind the Hydro distributed-systems framework says code review cannot keep up with agents. The team's answer is checks a compiler can enforce. Shadaj Laddad, the Hydro research lead at AWS, told The New Stack this at AGNTCon and MCPCon last week. Laddad calls the compiler layer "outside the agent's zone of trust."

Hydro is an open-source Rust framework for distributed systems. The software runs across many machines at once. Agents now write most of Hydro's own code. Humans still drive. Laddad told The New Stack the team recently finished a 5000-line change. The change checks a distributed system still works when it resends a message after a failure. The parts worked together, so the pull request was hard to split.

Why reviewing the code is not enough

Laddad told The New Stack that agents write both the code and the unit tests. "Agents are great, but they're unpredictable," he said. That unpredictability makes them creative. It also makes it hard for a person to check every test and confirm it does what the team wants.

A new feature can also pass alone and break in combination. Laddad told The New Stack a feature can work well in isolation. Mix it with something else, and the guarantees can break down.

The team's answer is to move the check to a layer the agent cannot write around. Rust's borrow checker is Laddad's example. The borrow checker catches race conditions before the code runs. Hydro does something similar for distributed systems. Several classes of bug are caught at compile time. The agent gets fast feedback on mistakes it introduced.

Shared guardrails are the goal

The team is studying historical projects for recurring bugs. The aim is to find bugs that could become general compile-time guardrails. One candidate, Laddad told The New Stack, is idempotency. Idempotency means a retried request does not run twice. The team sees it in design reviews often. The question is whether a compiler can enforce it for every project.

Compilers have limits. As The New Stack reported in August, AI-generated Rust can compile and still be wrong. Laddad frames this as the next rung on an old ladder. C spared programmers from mangling function calls in assembly. Java took memory allocation off the plate. Agents, he argued, make the next layer more urgent.

Infinity suspends an idle agent

The same team released Infinity on 23 October under Apache 2.0, as part of Hydro. Infinity is an agent runtime. It suspends an agent while the agent waits on a tool call. The same holds when the agent waits on a CI job. The agent is not kept running during the wait. Laddad cited a benchmark showing 75,000 concurrent agents on a single Raspberry Pi 5, as reported by The New Stack.

The AWS blog states this figure is extrapolated. It comes from a server test. Each idle agent used about 100 KB of memory in that test. Hydro and Infinity are at hydro.run, built jointly by AWS and UC Berkeley.

Hydro itself is an early release. Laddad told The New Stack users should expect rough edges. The payoff, in his words, is confidence to change a system. A team can change it without fearing it will break a few months later.

Source

The New Stack, Matthew Burns. AWS Marketplace sponsored the post. Project home: hydro.run. Infinity announcement: AWS Open Source Blog, Shadaj Laddad.