In a shift away from monolithic LLM prompts, developer Jules Langlois has proposed a modular approach to AI agent orchestration. On October 17, 2026, Langlois submitted PR #683 to the Neuron AI framework, introducing a 'Jarvis' agent mesh built entirely in PHP. The prototype, found in examples/agent/jarvis-mesh.php, demonstrates how small, role-specific agents can cooperate through a workflow engine rather than direct invocation.

Decoupling Agents via Typed Events

The core architecture rejects the trend of massive, all-knowing agents. Instead, it utilizes three distinct roles: Research, Operations, and Briefing. These agents never call each other directly. Instead, a Neuron Workflow routes execution using typed events such as BriefingStarted, ResearchCompleted, OperationsCompleted, and BriefingComposed. This design ensures the business logic graph is readable directly from type signatures, maintaining strict separation of concerns between agent capabilities and process flow.

Adapters for Transport and State

Langlois deliberately excludes scheduling and transport logic from the framework itself. The implementation relies on four custom interfaces: JarvisSignalBus, JarvisKnowledgeStore, JarvisHeartbeatStore, and JarvisAgentFactory. These adapters are passed to the workflow via a WorkflowResources object. This structure allows developers to swap LLM providers or storage backends without modifying the core workflow code, keeping the orchestration layer agnostic to infrastructure details.

Idempotency and Watchdog Mechanics

Resilience is handled through idempotent signal publication and memoization. Agent calls are memoized to prevent duplicate execution when a workflow resumes after an interruption. Signals use stable keys like '{workflowId}/research.completed' to ensure repeated publications do not create duplicates. A secondary WatchdogWorkflow monitors agent health by checking heartbeats against a staleness threshold. This watchdog is capped at one step with a 60-second lease, preventing infinite loops while alerting on stale or missing agents.

Key Takeaways

  • The prototype uses Neuron 4.1 and is currently unmerged, serving as a design pattern rather than production-ready code.
  • Workflow signals in Neuron are not broadcast queues; a separate idempotent bus is required for fan-out coordination.
  • Testing was performed with simulated agents and in-memory adapters, confirming that failed steps resume without repeating completed work.
  • External side effects remain the developer's responsibility to make exactly-once, as memoization cannot guarantee this for all integrations.

The Bottom Line

This PR is a pragmatic blueprint for moving AI agents out of the 'giant prompt' era and into structured, observable workflows. By forcing developers to implement their own transport and state adapters, it prioritizes architectural clarity over convenience, which is exactly what production systems need.