Scale Signal

Google paired nested coding-agent loops with build, test, monitoring, and fleet automation for an ISA migration spanning 100K+ applications

October 21, 2025 · Google Cloud
ApparitionSelectionContextCoordinationEconomicsExecutionObservabilityVerification
Scale signal: The source describes a production migration spanning an environment of more than 100,000 applications, 38,156 analyzed monorepo commits and about 700K changed lines, with more than 30,000 applications already migrated to Arm. CogniPort's formal early benchmark contains 245 historical migration commits and reports roughly 30% successful automated repair from rolled-back states.
Evidence record

Source → Observed → Interpretation → Model implication

SOURCE

At Google, the future is multiarch; AI and automation are helping us get there

View source →
OBSERVED

Google states that its production environment contains more than 100,000 applications that increasingly need to compile and run on both x86 and Arm. More than 30,000 applications had already been migrated to Arm at the time of publication, and services including YouTube, Gmail, and BigQuery were serving production traffic on both architectures.

The migration team analyzed 38,156 commits from Google's monorepo covering about 700,000 changed lines. Early migration work emphasized tooling, test adaptation, code changes, build and release systems, and production rollout issues; later phases shifted heavily toward configuration and supporting-process changes as migration volume increased.

Google reports that coordinating directly with application owners for the full long tail would have been cost-prohibitive, so the migration relied on automation to reduce application-team involvement. Existing systems included Rosie for generating and shepherding large numbers of code-review changes, sanitizers and fuzzers for architecture-specific failures, and CHAMP for automated rollout and health monitoring with eviction of unhealthy Arm jobs.

CogniPort is an AI migration agent that responds to build and test failures. If a target does not build or a test fails, the agent attempts to modify files until the target succeeds or the agent gives up.

CogniPort contains three nested agentic loops. An outer orchestrator repeatedly invokes a build-fixer agent and a test-fixer agent. Each loop produces a reasoning step and tool invocation, executes the tool, and attaches the result back into agent context for the next iteration.

The build-fixer repeatedly builds a target and edits files until the build passes or it gives up. The test-fixer repeatedly runs a test and edits files until the test passes or it gives up, and may invoke the build-fixer when test compilation itself fails.

Google created a benchmark of 245 historical Code & Test Adaptation commits that had originally been produced without AI assistance and could be cleanly rolled back. CogniPort was then evaluated on whether it could reproduce a successful fix from the rolled-back state.

Google reports that, without special prompts or other optimizations, the early benchmark successfully fixed the rolled-back failures about 30% of the time. The agent was particularly effective on test fixes, platform-specific conditionals, and data-representation fixes.

Google states that all new applications are now designed to be multiarch by default, while tens of thousands of older applications remain to be addressed through automation and application-owner involvement for more difficult cases.

INTERPRETATION

CogniPort is a bounded agentic engineering actor rather than an interactive coding assistant: it receives build and test failures, invokes tools, edits repository files, observes results, and iterates without continuous human operation until a verification condition is met or the run terminates. Its viability is inseparable from the surrounding engineering environment. Repository visibility supplies context; build and test systems define executable feedback; the orchestrator coordinates specialized repair loops; CHAMP and other automation provide observability and rollout control; and the economics of coordinating across more than 100,000 applications motivates automation in the first place. The 30% early benchmark result also exposes a substantial selection boundary: this configuration is useful for some migration classes but not yet reliable for the full long tail.

MODEL IMPLICATION

SUPPORTS. The source supports C01 by documenting an engineering configuration that can take repository-level action through iterative tool use and verification without continuous human operation. It supports C02 because the reported actor's fitness depends directly on Context, Execution, Verification, Coordination, Observability, and Economics conditions supplied by Google's engineering system. The evidence also shows differential effectiveness by migration problem type and a large unresolved long tail, indicating that the environment selects where this agent configuration remains viable rather than making it uniformly capable.

Epistemic boundaries

What this does not establish

OPEN QUESTION

As repository-scale migration agents move from selected repair classes into the remaining long tail, which combinations of context, verification, orchestration, rollout observability, and human handoff most improve successful fixes per unit of inference, review, and application-owner attention?