A joint paper from Google Research and MIT dropped this week with something the multi-agent community has been missing: principled scaling laws for multi-agent systems. No, not benchmarks or vibes. Actual mathematical frameworks for predicting how agent teams behave as you add complexity. If you're building systems with more than one agent, this paper is required reading.
The timing matters because the rest of the stack is crystallizing at the same speed.
The Merge Problem Gets Solved
Weave just shipped a semantic merge driver designed specifically for multi-agent codebases. If you've ever had two agents edit the same file and produce a git conflict that makes no semantic sense, you understand why this matters. Traditional three-way merge was designed for humans working on separate features. Agents don't work that way — they make holistic changes across files, and the conflicts they produce are fundamentally different from human conflicts.
Weave's approach treats merge at the semantic level rather than the text level. It understands what the code means, not just what it says. For teams running multiple coding agents in parallel — which is rapidly becoming the default workflow — this removes one of the most painful friction points in the pipeline.
GPT-5.4 Collapses the Stack
OpenAI's GPT-5.4 release collapses reasoning, coding, and computer-use into a single model with a one-million token context window and a new Tool Search API. The architectural implication is significant: some of the multi-model orchestration complexity that teams have been engineering around — routing different tasks to specialized models — can now potentially be handled at the model layer.
This doesn't eliminate multi-agent architectures. It changes what the agents are for. Instead of orchestrating across model capabilities, the orchestration layer moves up to workflow coordination, tool selection, and cross-system integration. The agents get simpler individually but the systems they compose get more ambitious.
Seatless Orchestration
Tess AI is pushing a "seatless" agent orchestration model — agents that don't require per-user licensing. This sounds like a pricing detail, but it's actually an architectural statement. When agents aren't tied to user seats, they can be deployed as infrastructure components rather than user tools. Think background processes that monitor, maintain, and optimize systems without human initiation.
What This Means
The multi-agent stack is moving from research artifact to engineering infrastructure. Scaling principles from Google/MIT, semantic merge from Weave, collapsed model capabilities from GPT-5.4, and infrastructure-grade orchestration from Tess — these aren't isolated developments. They're layers of a stack that's rapidly becoming buildable.
Six months ago, multi-agent systems were experimental. Today, the tooling exists to build them reliably. The question has shifted from "can we?" to "what's the right architecture?" — and that's an engineering question with engineering answers.
Happy thinking, Jason