Salesforce · Filed Jan 21, 2025 · Published Jul 23, 2026 · verified — real USPTO data

Salesforce Patents a System That Shifts AI Model Weights Mid-Conversation

Most AI systems run with a fixed personality from start to finish. Salesforce is patenting a way to slip in a different set of internal settings the moment one AI hands a message to another, without the user noticing a thing.

Salesforce Patent: Dynamic LLM Weight Switching Between Agents — figure from US 2026/0212117 A1
Figure from the official USPTO publication.
See all 8 drawings from this filing ↓
Publication number US 2026/0212117 A1
Applicant Salesforce, Inc.
Filing date Jan 21, 2025
Publication date Jul 23, 2026
Inventors Akash Singh, Rajdeep Dua
CPC classification 704/9
Grant likelihood Medium
Examiner THOMAS-HOMESCU, ANNE L (Art Unit 2656)
Status Docketed New Case - Ready for Examination (Feb 26, 2025)
Document 20 claims

How Salesforce's mid-conversation AI retuning works

Imagine you call a company's support line and the representative secretly gets a different script the second they transfer you to billing. The rep sounds the same, but their instructions have changed to match the new context. That's roughly what Salesforce is building for AI agents.

The patent describes a system with three AI agents working together. The first agent generates a response. A third agent, running silently in the background with the same starting instructions, reads that response and decides whether a setting change is needed. If it does, the underlying AI model swaps to a different set of internal tuning values before passing the message to a second agent.

Once the handoff is done, the model flips back to its original settings. The whole cycle happens automatically, based on the content of the message itself, not a manual instruction from a human.

How the three-agent handoff triggers a weight swap

The patent covers a method for adjusting what engineers call parameter weights (the numerical values inside an AI model that shape how it responds) dynamically, based on what the AI just said.

The setup uses three coordinated AI agents:

  • Agent 1 generates a message using a primary prompt and its associated settings.
  • Agent 3 is a silent monitor, configured with the same initial instructions as Agent 1. It reads Agent 1's output and produces a control message, essentially a judgment call about that content.
  • Agent 2 is the downstream recipient, running under a different prompt and different parameter set.

When Agent 3's control message signals that a change is warranted, the system switches from the model's first set of weights to a second set before forwarding Agent 1's message to Agent 2. After delivery, it switches back to the original weights.

The practical effect is that the AI model can behave differently depending on what it just produced, without requiring a human to intervene or a completely separate model to be loaded. The weight swap is temporary and content-driven.

We find one patent like this every day. Get the best of each week in your inbox, free →

What this means for AI agents handling complex tasks

Multi-agent AI pipelines are increasingly how enterprise software (including Salesforce's own Agentforce platform) handles complex workflows: one AI drafts, another reviews, another takes action. If each agent is locked into a single behavioral profile, the chain can produce awkward handoffs where the tone, format, or detail level is mismatched.

This patent suggests Salesforce wants the model itself to adapt mid-pipeline, not just the prompt. For enterprise customers, that could mean AI workflows that handle a customer complaint differently than a routine data query, even when both start from the same agent template. The real question is how much overhead the weight-switching adds at scale, since swapping model parameters is not a cheap operation.

Editorial take

This is a genuinely interesting architectural idea for multi-agent AI systems, and it fits squarely into Salesforce's push to make Agentforce capable of nuanced, context-sensitive behavior. The three-agent monitor pattern is clever and sidesteps some of the bluntness of prompt-only conditioning. Whether it performs well enough in production to ship is a separate question, but this is worth watching.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings

8 drawing sheets from US 2026/0212117 A1 · click any drawing to enlarge

Patent filing page

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

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