Gartner logged a 1,445% surge in multi-agent system inquiries between Q1 2024 and Q2 2025. That's not a typo. The number is absurd enough that it tells you something about where corporate attention has landed, even if it tells you very little about whether anyone has actually figured this out.

They haven't.

Full agent orchestration — where multiple specialised AI agents coordinate autonomously on complex tasks, handing off context, negotiating subtasks, recovering from failures without human intervention — remains aspirational. The pieces exist. The plumbing is getting built. But the thing itself, the seamless multi-agent workflow that enterprise slide decks keep promising, isn't here yet. Not in any form I'd trust with real work.

Here's where things actually stand. GitHub launched Agent HQ this week with Claude, Codex, and Copilot all available as coding agents. You can assign different agents to different tasks from issues, pull requests, even your phone. Anthropic's Claude Agent SDK supports subagents that spin up in parallel, each with isolated context windows, reporting back to an orchestrator. The infrastructure for coordinated work is plainly being assembled. I wrote about this trajectory a week ago — the session teleportation, the hooks system, the subagent architecture all pointing toward something more ambitious. That trajectory has only accelerated.

The gap between "agents that can be orchestrated" and "agents that orchestrate themselves" is enormous, though. And it's not a gap that better models alone will close.

Consider the context problem. When you connect multiple MCP servers — which is how agents typically access external tools — the tool definitions and results can bloat to hundreds of thousands of tokens before the agent even starts working. Anthropic's own solution compresses 150K tokens down to 2K using code execution sandboxes, which is clever, but it's a workaround for a structural problem. Orchestrating multiple agents means multiplying this overhead across every participant. The economics don't hold up yet.

Then there's governance. Salesforce's connectivity report found that 50% of existing agents operate in isolated silos — disconnected from each other, duplicating work, creating what they diplomatically call "shadow AI." 86% of IT leaders worry that agents will introduce more complexity than value without proper integration. These aren't hypothetical concerns. The average enterprise runs 957 applications with only 27% of them actually connected to each other. Drop autonomous agents into that landscape and you get chaos with better branding.

Security is the other wall. Three vulnerabilities in Anthropic's own Git MCP server enabled remote code execution via prompt injection. Lookalike tools that silently replace trusted ones. Data exfiltration through combined tool permissions. These are the kinds of problems that get worse, not better, when you add more agents with more autonomy. An orchestrator coordinating five agents is also coordinating five attack surfaces.

I spent the last week building a video generation app that uses four different AI models through the same interface. Even that simple form of coordination — one human choosing which model to invoke, with no inter-agent communication at all — required model-specific API contracts, different parameter schemas, different pricing structures, different prompt styles. One model wants duration as "8", another wants "8s". One supports audio, another doesn't. Multiply that friction by actual autonomy and you start to see why this is hard.

So how long? My honest guess: we'll see convincing demonstrations of multi-agent orchestration in controlled environments within the next six to twelve months. GitHub Agent HQ is already close for the narrow case of software development. The patterns are converging — Anthropic's subagent architecture, MCP as the connectivity standard, API-centric integration layers. Deloitte projects that 40% of enterprise applications will embed task-specific agents by end of 2026.

But "embed task-specific agents" is not the same as "full orchestration." Embedding a specialised agent into a workflow is plugging in a power tool. Full orchestration is the tools building the house while you sleep. We're firmly in the power-tool phase, and the industry keeps selling blueprints for the house.

The honest answer is probably two to three years for production-grade, genuinely autonomous multi-agent orchestration in enterprise settings. And that assumes the governance and security problems get solved in parallel with the technical ones, which — given how security usually goes — feels optimistic. The models are ready. The protocols are converging. The trust isn't there yet, and trust is the bottleneck that no amount of architectural cleverness can route around.

Sources: