Published and updated 7 September 2026 · Altari Systems

Distributed execution

Fenced Execution for AI Coding Agents

Abstract

Fenced execution gives each coding-agent assignment a lease and a monotonically increasing ownership token. A worker must present the current token when it delivers. If the lease expires and work is reassigned, the old worker's later result is rejected even if that worker never learned it had lost ownership.

01

Cancellation messages are not enough

A worker can lose network access, pause or ignore cancellation while continuing to edit. The scheduler cannot assume that sending stop made the old process harmless.

The receiving authority must validate ownership at the moment a result tries to change project state.

  • Lease expiry
  • Monotonic generation
  • Delivery-time validation
  • Idempotent transition
02

Fencing protects against late success

The dangerous worker is not only one that fails. A previous worker may eventually succeed after a replacement has produced newer work.

Rejecting the stale token prevents both results from claiming current authority. The old evidence can remain available for diagnosis.

  • Timeout and reassignment
  • Partition recovery
  • Duplicate delivery
  • Scheduler restart
03

Nexus implementation context

Nexus applies fenced ownership to bounded software slices stored in durable state. Repository candidates and evidence must correspond to current authority before review advances.

This is a core mechanism for safely running multiple AI coding agents, not a user-interface convention.

Nexus launch access

Be first to know when Nexus launches.

Register your interest now and we will let you know when the autonomous AI software engineering control plane is ready.

Express Interest ↗