Google · Filed Nov 27, 2024 · Published May 28, 2026 · verified — real USPTO data

Google Patents a Multi-Agent System That Merges Conflicting AI Prompts

What happens when two AI agents disagree about how to phrase a request to a generative model? Google's latest patent describes a system that resolves those conflicts automatically, merging multiple agents' prompts into one coherent request before anything gets sent to the model.

Google Patent: Multi-Agent AI Prompt Collaboration System — figure from US 2026/0148012 A1
FIG. 1A — rendered from the official USPTO publication PDF.
Publication number US 2026/0148012 A1
Applicant GOOGLE LLC
Filing date Nov 27, 2024
Publication date May 28, 2026
Inventors Ajay Prasad
CPC classification 704/9
Grant likelihood Medium
Examiner RILEY, MARCUS T (Art Unit 2654)
Status Docketed New Case - Ready for Examination (Jan 3, 2025)
Document 20 claims

What Google's collaborative prompt-merging system actually does

Imagine a team of coworkers all trying to send instructions to the same assistant at the same time — but their instructions contradict each other. One person says 'use Canny edge detection,' another says 'use a different algorithm entirely.' Without someone to arbitrate, the assistant gets confused.

Google's patent tackles exactly this problem, but for AI agents — software programs that each generate their own natural language prompt to send to a generative model like Gemini. Instead of letting all those prompts pile up or collide, the system aggregates them into a single unified prompt, detecting and resolving conflicts along the way.

The key step is a disambiguation process: when two agents propose conflicting parameters, the system picks one — based on rules or context — and builds the final, combined prompt around that decision. The result is one clean, coherent request that the generative model can actually work with.

How the disambiguation process resolves conflicting parameters

The patent describes a server-side coordination layer sitting between a plurality of agents (multiple independent software processes, each producing its own natural language prompt) and a single generative model (GM).

Each agent submits a prompt containing one or more candidate parameters — specific instructions or settings embedded in natural language, like which algorithm to use or what output format to produce. The system collects all of these prompts and attempts to combine them into a joint NL prompt.

The interesting part is conflict resolution. When two agents include incompatible candidate parameters — say, one wants Canny edge detection and another wants a different image-processing method — the system flags the clash and runs a disambiguation process (a rule-based or model-driven decision procedure) to pick one parameter over the other. That selection is then factored into the final merged prompt.

  • Step 1: Collect individual NL prompts from each agent
  • Step 2: Identify conflicting parameters across prompts
  • Step 3: Run disambiguation to select the winning parameter
  • Step 4: Construct the joint NL prompt incorporating that decision
  • Step 5: Send the joint prompt to the generative model for a response

The abstract specifically calls out a generative coding application as an example use case, suggesting this is designed partly for agentic coding workflows.

What this means for multi-agent AI pipelines

As AI pipelines get more complex, you're increasingly dealing with multiple agents operating in parallel — each with its own sub-goals, preferences, and prompt strategies. Sending all of those to a model independently is wasteful and inconsistent; merging them naively produces incoherent garbage. A coordination layer that resolves conflicts before the model ever sees the prompt is a real engineering need, not a theoretical one.

For Google, this fits squarely into its push to make Gemini the backbone of multi-agent developer tooling. If agentic coding assistants, research tools, and workflow automations all need to share a single model endpoint, something has to manage the traffic — and this patent describes exactly that arbitration logic.

Editorial take

This is a genuinely useful piece of infrastructure for multi-agent AI systems, and Google filing it now makes sense given where agentic AI tooling is headed. The core idea — conflict-aware prompt merging — solves a real coordination problem that anyone building multi-agent pipelines has already hit. It's not flashy, but it's the kind of plumbing that makes complex systems actually work.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.