PromisingAgents & OrchestrationMoved upMarch 2026

Strong signal and real results. Worth committing a pilot to.

Multi-agent Orchestration

Multi-agent systems moved from experiment to production pattern — Gartner predicts 40% of enterprise apps will embed AI agents by end 2026.

Agentic·Infrastructure

langchain.com

Our Take

What It Is

Multi-agent orchestration is the practice of coordinating multiple AI agents to collaborate on complex tasks. Rather than one model doing everything, you decompose work across specialised agents: a planner, a researcher, a coder, a reviewer. Frameworks include LangGraph, CrewAI, Microsoft Autogen, and the emerging A2A protocol for cross-framework agent communication.

Why It Matters

We moved multi-agent to Promising because the production evidence now outweighs the research hype. Gartner's 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025 reflects genuine enterprise demand, and their prediction that 40% of enterprise apps will embed AI agents by end 2026 (up from 5% in 2025) sets a clear adoption trajectory.

The practical breakthrough is the Plan-and-Execute pattern: use a frontier model (Claude Opus, GPT-5.4) as the planner and cheaper models (Haiku, GPT-5.3 Instant) as executors. Teams report 90% cost reduction compared to running everything on a single frontier model. VS Code 1.109-1.110 adding native multi-agent support signals that the tooling is catching up to the pattern.

Key Developments

  • Mar 2026: Google's A2A protocol donated to Linux Foundation with 50+ partners including Salesforce, SAP, and ServiceNow.
  • Feb 2026: VS Code 1.109-1.110 adds multi-agent orchestration support with inter-agent communication.
  • Jan 2026: Gartner predicts 40% of enterprise apps embed AI agents by end 2026. Market projected $7.8B to $52B by 2030.
  • Dec 2025: Plan-and-Execute pattern (frontier planner + cheap executors) emerges as dominant cost-optimisation approach.

What to Watch

The interaction between MCP (agent-to-tool) and A2A (agent-to-agent) will define the orchestration architecture for the next two years. Watch for whether teams can actually debug multi-agent systems in production — observability across agent chains is the current weak point. If Langfuse or Braintrust ship first-class multi-agent tracing, that removes the biggest practical blocker for enterprise adoption.

Strengths

  • Cost efficiency: Plan-and-Execute pattern with frontier planner + cheap executors reduces costs up to 90% compared to single-model approaches.
  • Task decomposition: Complex workflows benefit from specialised agents — a coding agent doesn't need to be great at research, and vice versa.
  • Standards forming: A2A protocol and MCP together provide a complete connectivity layer for interoperable agent systems.
  • Enterprise demand: Gartner's 40% enterprise agent prediction and 1,445% inquiry surge indicate real market pull, not hype.

Considerations

  • Debugging complexity: When a multi-agent pipeline fails, tracing the failure across agents is significantly harder than debugging a single model call.
  • Latency accumulation: Each agent hop adds latency. Multi-step orchestrations can be noticeably slower than monolithic approaches.
  • Error amplification: A misinterpretation by one agent can cascade through the pipeline, with each subsequent agent amplifying the original error.
  • Operational overhead: Managing multiple agent configurations, prompts, and model selections multiplies the operational surface area.