Definition

Multi-agent architectural drift is what happens when multiple autonomous coding agents make locally reasonable changes that gradually diverge from shared architectural decisions, abstractions, and repository conventions.

Why parallel agents drift

Each agent acts on the context it has. None of them has the full picture of what the others are doing in parallel, what conventions a senior engineer would invoke, or which patterns have been deprecated in conversations that never made it into the repo.

Without a shared enforcement layer, the agents do not need to be wrong individually for the system to drift collectively. Each change is explainable in isolation. The aggregate is not.

An example

Three agents work on the same codebase in parallel:

  • One adds a direct database dependency to a service it was modifying.
  • Another creates a parallel service abstraction because the existing one looked complicated.
  • A third copies a deprecated pattern because it exists nearby in the codebase.

Each change is locally reasonable. Together, they weaken the architecture. No single PR review catches the pattern — each looks fine on its own.

Why review does not scale

Human reviewers are sequential. Autonomous agents are parallel. The throughput mismatch is the failure mode: by the time the third PR lands, the first one’s precedent has already been set, and the second one has already justified its own deviation by pointing at the first.

PR review can still catch egregious violations. It cannot maintain coherence across a stream of plausible ones.

Why memory is insufficient

Memory systems — embeddings, retrieval, long context — can surface the relevant ADR or convention to the agent. They cannot enforce that the agent followed it. Recall is not constraint. The agent might be told about the rule and choose a different path because the local task seemed to call for it.

Memory makes the rule available. Governance makes the rule binding.

Why deterministic governance is needed

The right primitive is a binary verdict: given the architectural corpus and the proposed change, the system says allowed or not allowed. Same input, same verdict, every time. That property is what lets the same constraint hold across parallel agents that never see each other’s context.

How Mneme reduces drift

Mneme compiles ADRs and architectural decisions into a deterministic corpus that every agent and CI surface evaluates against. The same compiled rules reach every agent in the swarm. Verdicts are reproducible. Provenance traces back to the originating decision. Parallel agents stop being able to drift in incompatible directions because the boundary is the same for all of them.

Related concepts