This tracker follows patents on routing tasks to the right specialist AI, merging conflicting instructions, and catching AI mistakes before they reach the user. Together, the filings show Big Tech betting that coordinating specialized models beats relying on one model to do everything.
108 filings · tracking since May 2026 · latest Sep 2026 · updates weekly
based on all tracked filings in this watchlist · refreshes every week
These filings are all fighting over the same core idea: instead of one AI doing everything, you build teams of AIs that check each other, hand off tasks, and fix their own mistakes before anyone sees them.
Microsoft leads by raw filing count, with Google close behind and a strong late push that keeps the race tight.
a dated entry each week this watchlist moves · older entries stay archived
This week's filings center on routing: getting the right question to the right AI tool or model automatically. Nvidia and Google led the pack, each filing multiple patents around directing tasks, splitting work, and picking the best AI for the job.
This week's filings lean heavily on AI models checking each other's work, with Google, Microsoft, and Samsung each filing multiple patents in that space. Microsoft and Samsung are the most active, each adding three filings to the watchlist.
Most new filings this week come from Microsoft, covering AI systems that fix their own errors, share memory across models, and build or recruit other AI agents as needed. The overall theme is AI models that check, correct, and assist each other rather than working alone.
This week's filings lean heavily toward AI systems that split work across multiple models, sending each task to whichever model handles it best. Nvidia and Qualcomm each filed twice, covering ideas from home network task routing to saving data only when two AI models disagree.
This week's filings lean heavily toward AI systems that check and correct their own work before giving an answer, with Samsung leading the pack with two separate filings on that idea. Salesforce and IBM each added one filing exploring similar self-checking behavior from different angles.
counts from tracked filings · focus read from each company’s own filings
20 or more filings in the last 8 weeks · 6 to 19 · under 6
the fights inside the fight · each with its three newest filings · new filings join every week
Salesforce 5, Google 4, IBM 4
Several companies are filing patents where one AI checks, scores, or rewrites its own output before a person ever sees it. Google, Salesforce, IBM, Nvidia, and Samsung are all pushing versions of this idea.
Google 6, Nvidia 4, IBM 4
Rather than one AI doing everything, these patents describe a system that reads the task and hands it to whichever specialist model fits best. Google, IBM, Qualcomm, and Microsoft are all filing in this space.
Microsoft 4, Google 3, Qualcomm 2
These patents pair two AI models so one produces an answer and the other checks or improves it. Google, Microsoft, Salesforce, IBM, and Qualcomm each have filings describing this two-model setup.
Microsoft 8, Salesforce 4, Nvidia 2
These filings describe networks of AI agents that split up a big task, pass results between each other, and coordinate so no two agents talk over each other. Microsoft, Salesforce, Nvidia, and Google are the main filers here.
Microsoft 3, Adobe 2, Google 2
These patents tackle the moment a person gives an AI a confusing or unclear request, with one model rewriting or clarifying the prompt before any real work begins. Adobe, Google, and Salesforce are the main filers.
Samsung 3, Qualcomm 2, IBM 1
Instead of locking into one AI model for a whole job, these patents let a system switch to a different model when conditions change or a better fit is available. Samsung, Qualcomm, IBM, and Microsoft are filing here.
AI agent teams that build missing members on demand could change how software gets made.
Switching between AI assistants mid-conversation with one word is a very consumer-visible idea.
An AI assistant that auditions answers before speaking could mean far fewer wrong replies.
A two-model AI that grades its own image drafts before showing them to you is notable.
every tracked filing, month by month · counts are USPTO pre-grant publications, one per publication number (an application, not a granted patent)
US 2026/0278268 A1
Users could deploy multiple AI agents without worrying that one will leak confidential data to another by accident. The filing shows how to enforce access controls at every handoff between agents, preventing information from crossing departmental boundaries.
US 2026/0277977 A1
After routing tasks to specialists, the next friction point emerges: those specialists need better instructions. This filing shows how one AI can iteratively improve prompts for another, automating what's now manual tuning.
US 2026/0277949 A1
Coordinator AI that farms specialist models and merges their answers, rather than one generalist model. Confirms the routing and synthesis piece of the team approach.
US 2026/0278453 A1
Users wouldn't need to manually test models against each task anymore. IBM's approach automates the selection process entirely, running the evaluation and scoring that typically requires human judgment or expensive experimentation.
US 2026/0278390 A1
After routing and conflict resolution, this filing adds the selection layer: how a dispatcher AI actually picks which specialist handles each subtask, moving from static task division to dynamic agent matching.
US 2026/0278290 A1
After routing and merging instructions, the next bottleneck is context. Google's filing shows how to pull relevant documents into the AI's working memory before it executes the user's chosen task, reducing the chance of generic or off-target outputs.
US 2026/0278349 A1
A secondary AI monitors the main chatbot's output and rewrites failed prompts without user intervention, removing the need for manual rephrasing cycles.
US 2026/0277950 A1
After routing tasks to specialists, this filing shows how to let agent teams build their own knowledge base from raw documents, then verify answers against it before users see them.
US 2026/0279336 A1
Splitting audio generation into separate models for content selection and delivery lets each specialize without the overpolite artifacts that mark machine speech.
US 2026/0277971 A1
A front-end dispatch layer that routes queries to specialized models without user intervention, solving the efficiency problem when multiple trained systems run in parallel.
US 2026/0277967 A1
After routing questions to specialist tools, this filing shows how to keep those tools from making independent errors, it describes a verification layer that checks whether the routed tool actually has reliable access to the data it claims to answer from.
US 2026/0268207 A1
Matching model capacity to hardware constraints and cost limits requires real-time inventory of available cloud resources. Microsoft's filing automates that matching step.
US 2026/0267727 A1
If one model flags an anomaly the other doesn't, the system alerts humans instead of silently missing it. Sony's patent shows how parallel models with different context catch edge cases a single detector would overlook.
US 2026/0268163 A1
Combining specialized models without retraining confirms the merge-and-weight approach: rather than pick one expert or train from scratch, blend trained models at specific ratios to preserve each model's strengths in a single output.
US 2026/0268890 A1
For multi-agent workflows to feel seamless to users, someone has to orchestrate which AI does what and when. Nvidia's filing shows how to hand off from script-writing to character assignment to voice generation without human judgment calls in between.
US 2026/0264721 A1
Catching AI errors mid-execution: Samsung's system routes autonomous driving decisions through a language model for safety review before the car acts, rather than only checking outputs after mistakes occur.
US 2026/0268214 A1
Adds a second model that can actually see when a primary model is optimizing for user approval rather than accuracy, solving the deception risk that training alone won't catch.
US 2026/0267900 A1
Verification layers between the user and the model output would catch confabulated facts before delivery, shifting the error-catching work from human review to automated checks.
US 2026/0268699 A1
Generating coherent multi-sentence captions requires catching hallucinations before output. Microsoft's three-model approach separates draft generation from fact validation, reducing false details in longer descriptions.
US 2026/0270224 A1
Coordinating specialist agents on sequential ad tasks (brief to score to assemble) without human review loops between steps.
US 2026/0267636 A1
Verifying code translation requires a second model to audit the first one's work rather than trusting a single AI's output, reducing the risk of silent errors when converting large codebases between languages.
US 2026/0260065 A1
Reducing latency in token generation. Microsoft's approach uses a smaller draft model to propose text in bulk, then a larger verifier checks the batch at once instead of validating each token individually, cutting redundant compute steps.
US 2026/0259722 A1
Catching build errors mid-process lets one AI fix failures without halting the entire sequence and escalating to human developers.
US 2026/0260061 A1
After routing work to specialist AIs, you need them to follow instructions correctly. This patent covers automating the fix when they don't, by having the AI diagnose its own mistakes and rewrite its own prompts.
US 2026/0260391 A1
Users could stop regenerating entire images when minor details fail: the system identifies which parts need work and refines only those sections instead of starting over.
US 2026/0259708 A1
The team-routing tracker so far covers delegating work to specialists and catching errors before users see them. This filing shows how one specialist might self-correct mid-task, salvaging working code fragments instead of restarting entirely.
US 2026/0259924 A1
After routing and merging, the next vulnerability is discovering mid-task that no specialist exists for a subtask. This filing shows how to auto-generate and integrate a new AI agent into the team without stopping work.
US 2026/0260148 A1
A routing layer that picks the right mathematical solver for each problem type, cutting wasted compute cycles when the wrong algorithm gets assigned by default.
US 2026/0260076 A1
If it works, users get responses matched to what they actually need instead of generic answers. Microsoft's approach splits the job between one AI that interprets the problem and another that generates the reply.
US 2026/0260652 A1
Previous filings show companies routing tasks between specialist AIs; this one reveals how a single conversation can swap which AI is answering without the user detecting the handoff, letting companies patch weak spots without full retraining.
US 2026/0260478 A1
After routing tasks to specialist AIs, this filing shows how to split visual analysis itself: one model plans the reasoning strategy while another executes the detailed looking, letting each focus on what it does best.
US 2026/0259830 A1
After establishing how to route work between specialist models, this filing cuts the redundancy within handoffs: instead of each model rebuilding its understanding of a conversation from scratch, one model's extracted context passes directly to the next one.
US 2026/0253296 A1
Users in mixed human-AI teams would see which AI is working on what task in real time, solving the opacity problem that arises when multiple specialists operate in the same workspace.
US 2026/0253256 A1
Letting users swap image engines mid-stream solves a routing problem that fixed pipelines can't: different vision models excel at different photo tasks, so locking into one wastes capability on the device.
US 2026/0252317 A1
Users get working code without manual debugging if an AI can test its own output and iterate until tests pass, solving the verification problem within a single specialized agent.
US 2026/0252974 A1
Routing new tasks to existing specialist models instead of retraining a generalist cuts the data and compute needed to add capabilities. Confirms the bet that multi-agent systems beat single large models for efficiency.
US 2026/0252401 A1
Deciding which local device handles an AI task based on capability and load. Nvidia's dispatcher lets home networks offload to idle hardware instead of running everything locally or routing to distant servers.
US 2026/0253383 A1
Detecting when parallel models reach different conclusions lets systems flag uncertain moments instead of storing continuous streams. This captures the routing problem at its source: deciding which cases need human review or escalation to a specialist model.
US 2026/0244680 A1
A verification loop where one AI model questions another's visual counts catches arithmetic errors before they reach users, solving a specific failure mode in image analysis that trips up even advanced systems.
US 2026/0245698 A1
The watchlist has focused on routing tasks between specialists. This filing shows the inverse problem: how to identify which specialist deserves more work by observing which one users actually follow.
US 2026/0244660 A1
The routing problem gets sharper: Samsung's system scores answers from multiple sources before committing to one, catching bad matches that single-path assistants would miss.
US 2026/0244996 A1
The routing problem gets a guard rail: IBM's patent lets a degraded team self-assess before answering, blocking silent failures when models drop offline.
US 2026/0239067 A1
Keeping AI error-checking local to the device instead of relying on network feedback loops. The proxy model catches mistakes on the phone's wireless chip itself, removing latency from the verification cycle.
US 2026/0239027 A1
Running multiple AI models on a device creates a bottleneck deciding which one to use when. Qualcomm's solution routes based on real-time signal conditions, moving the selection logic off the main processor to the wireless chip itself.
US 2026/0237196 A1
Routing through a fast screening model before deploying expensive compute avoids wasting resources on irrelevant suggestions when the stakes are high but the user's intent is unclear.
US 2026/0236461 A1
Semantic parsing errors in plain-English queries need detection before output. Adobe's system automatically identifies which command types an AI repeatedly misinterprets, enabling targeted fixes that prevent bad results from reaching users.
US 2026/0236586 A1
Catching design flaws before fabrication requires AI agents that can parse specialized code and spot errors humans might miss. Microsoft's system routes chip-programming tasks to specialized reviewers, automating the multi-day review cycle.
US 2026/0236825 A1
Users could hand off spec-writing to AI pairs that draft and validate each other, cutting the manual labor of hunting contradictions across thousands of pages of accumulated requirements.
US 2026/0236235 A1
Generating synthetic training data requires custom code, and IBM's filing shows how one AI can produce multiple candidate generators while another ranks them by output quality, cutting the manual labor of writing and vetting data pipelines.
US 2026/0228047 A1
A broker system that spins up fresh agent instances on demand and tracks resource availability across servers, directly solving the coordination problem when dozens of specialists run simultaneously without colliding or competing for the same compute.
US 2026/0228506 A1
Inspecting the model's internal state rather than just parsing user requests lets the system route tasks with higher confidence, reducing mismatches between what users ask for and what skills the AI actually applies.
US 2026/0228504 A1
A secondary AI refines task prompts instead of retraining the primary model, cutting the computational cost of fixing mistakes by iterating on instructions rather than weights.
US 2026/0228495 A1
Sharing a common intermediate layer among specialist models reduces the parameter count that multi-expert systems must load and compute, lowering the overhead when routing tasks across the team.
US 2026/0228513 A1
Routing uncertain outputs to a verification model instead of running everything through the larger system cuts computation while keeping quality intact, solving the cost problem that's blocked team-based AI in production.
US 2026/0228540 A1
Running multiple specialist modules in parallel and weighting their outputs by input type offers a hardware-efficient alternative to swapping between separate models, letting inference happen simultaneously rather than sequentially.
US 2026/0229236 A1
When routing between specialist AIs, Samsung solves the friction problem: users can switch assistants in real time without leaving the conversation, using voice commands instead of manual settings changes.
US 2026/0220130 A1
The routing watchlist needed a solution for questions that arrive malformed. Adobe's filing shows how to catch unclear user input before dispatch, letting a reformulation layer clarify intent so the right specialist AI gets workable instructions.
US 2026/0220131 A1
Routing uncertain queries to human experts rather than generating unreliable answers adds a feedback loop where the system learns from each escalation, strengthening the team's collective knowledge over time.
US 2026/0220028 A1
A separate AI evaluates whether test cases written by a code-generating AI actually expose bugs, adding a verification layer that catches when automated code-writers produce tests that look good but miss real failures.
US 2026/0221132 A1
Within the team watchlist, this filing shows how to keep a primary model in place while adding a specialist downstream that learns which errors to expect from it specifically, rather than retraining the whole system or swapping in different models.
US 2026/0220434 A1
An intermediate AI layer converts each model's native representations into a shared format, eliminating the text conversion bottleneck that slows multi-AI workflows and degrades information fidelity.
US 2026/0220536 A1
Automatically ranking models on unfamiliar data sidesteps the need for human labeling when routing tasks to the right specialist, letting systems pick their own best performer instead of waiting for engineers to validate each option.
US 2026/0222368 A1
The watchlist has focused on routing tasks to specialists. Sony adds a layer below that: making each specialist develop its own consistent voice while staying aligned with shared objectives.
US 2026/0220451 A1
The watchlist has focused on routing and merging. This filing adds a verification layer: a second model that evaluates the first model's outputs before they leave the system, catching errors in real time rather than after deployment.
US 2026/0220166 A1
A second AI grades the first one's work against quality criteria and bounces failing responses back for revision, keeping errors from reaching users without requiring human review of every output.
US 2026/0220214 A1
Where the watchlist has focused on routing tasks upfront, this filing shows a system that routes mid-process: a lightweight model runs first, then checks its own confidence to decide whether to hand off to a specialist model before returning results.
US 2026/0211976 A1
One AI model grades another's repeated attempts at the same question, creating paired examples of right and wrong answers that train the weaker model to recognize its own errors without human labeling.
US 2026/0212128 A1
Users could skip the handoff delays: one plain-English query routes to the right specialist AI automatically, pulling sensor data and prediction models together to answer maintenance questions without human intermediaries.
US 2026/0211913 A1
Splitting spatial reasoning between a language model and a specialized geometry agent lets teams handle queries that need both semantic understanding and precise coordinate math, a division that sidesteps trying to cram spatial logic into text-only systems.
US 2026/0212266 A1
The watchlist so far has focused on routing tasks and catching errors. This filing adds a middle layer: one AI agent that watches and modifies another agent's output in real time, reshaping both what gets said and how the next agent will receive it.
US 2026/0212235 A1
The routing watchlist so far assumes specialists stay ready in parallel. IBM's filing confirms the opposite works too: swapping which models run based on real-time passenger state and vehicle conditions cuts the compute needed at any moment.
US 2026/0212117 A1
The routing challenge deepens: keeping handoffs seamless when each specialist needs different internal weights to work properly on its assigned task.
US 2026/0211953 A1
The watchlist so far has focused on routing work to the right specialist. This filing shows the alternative approach: run multiple specialists in parallel and use selection logic to pick the strongest output, sidestepping the routing problem entirely.
US 2026/0203613 A1
Users get faster compliance checks when one AI reads regulations while another verifies systems actually meet them, rather than relying on manual legal review.
US 2026/0203643 A1
Users get faster answers when models skip dormant layers, cutting compute waste on routine queries. The filing confirms Big Tech is optimizing which internal components activate per request, not just which external specialist model to call.
US 2026/0203339 A1
Users get both a ranking and reasoning they can actually read. The filing shows how to split scoring from explanation so the justification doesn't just mirror the decision, it can point out what actually moved the needle.
US 2026/0203553 A1
Users get faster answers and lower costs when the system learns which specialist modules matter for each task, rather than running full computation every time.
US 2026/0203286 A1
Stopping hallucinations requires splitting the work: one model identifies what facts to retrieve, another generates answers from only those results, creating a hard checkpoint before output reaches users.
US 2026/0192824 A1
The routing watchlist now moves from task distribution to real-time decision arbitration: how to run competing planning systems in parallel and filter their outputs through a safety layer before execution.
US 2026/0195656 A1
An ensemble that chains specialist models to handle only the cases earlier models failed on, rather than retraining the whole system each time.
US 2026/0186828 A1
The routing problem needs memory: when specialist models hand off tasks, they currently lose what they learned. This filing adds a shared knowledge layer so downstream agents don't duplicate work or contradict earlier decisions.
US 2026/0189558 A1
Controlling what data flows between specialist agents in a multi-agent system prevents one poorly designed or compromised agent from leaking sensitive information to another that shouldn't access it.
US 2026/0187443 A1
The watchlist has focused on routing tasks and merging instructions; this filing shows how to verify outputs once models finish. Multiple detectors voting on sensitive data catches errors that single-path systems miss.
US 2026/0187482 A1
The routing layer needed a gatekeeper that doesn't just split traffic but learns which model handles which question type better over time. Nvidia's filing shows how to build that learned routing rather than relying on fixed rules.
US 2026/0186749 A1
The routing watchlist gains a concrete use case: automated code repair that matches error type to specialized models, showing how task assignment works when mistakes need real-time correction.
US 2026/0187127 A1
A cheaper filter screens video frames before routing them to the main model, cutting processing load when one specialist AI alone would waste capacity on irrelevant content.
US 2026/0187995 A1
A cascading sequence of specialized models, each refining the previous output, confirms the watchlist's assumption that video quality improves when tasks flow through separate AI stages rather than a single model handling everything at once.
US 2026/0186166 A1
Running physics equations and neural networks in parallel lets each method correct the other's blind spots, physics catches when AI extrapolates beyond its training data, while AI accelerates calculations that would otherwise consume too much compute.
US 2026/0187696 A1
Sequential handoff between specialist models shows how to keep output quality high when each AI handles one piece of a workflow, from selection through presentation.
US 2026/0189498 A1
A routing layer that tracks which agent has handled prior interactions with each customer, ensuring continuity instead of forcing each incoming message through a generic assignment process.
US 2026/0189557 A1
Multi-agent systems need ways to specify which model can access what data. Using natural language rules instead of code makes permission-setting faster and reduces bottlenecks when orchestrating specialist agents.
US 2026/0181031 A1
Coordinating across human-robot-AI teams requires merging separate perception streams into one shared workspace where all participants see the same operational picture.
US 2026/0178028 A1
Routing robot control through a shared workspace means teams can hand off between automation and human operation without losing sight of what others are doing.
US 2026/0178027 A1
The watchlist has focused on routing tasks between specialist models. This filing adds a layer before that: translating human intent into machine instructions that the routing system then executes.
US 2026/0170345 A1
Routing incoming queries to the right specialist model based on content analysis, rather than random assignment or sequential processing, solves the selection problem that emerges when teams of models replace single monolithic systems.
US 2026/0161743 A1
Validating AI outputs requires specialist judges rather than one generalist model, with a master evaluator weighing their separate verdicts on grammar, accuracy, and tone before approval.
US 2026/0161712 A1
A coordinator layer that routes to specialists based on question content could cut down on hallucination by keeping each model within its trained domain rather than forcing generalists to improvise across fields.
US 2026/0161893 A1
The routing challenge shifts here: instead of sending tasks to specialist models upfront, the system lets a generalist model draft first, then routes only the problematic sections to editors. This inverts the usual workflow.
US 2026/0154561 A1
The routing problem gets sharper: which specialist model should handle summarization depends not on summary quality alone, but on whether that summary actually serves the next model's task.
US 2026/0154534 A1
Scaling safety validation across modalities requires automated grading that doesn't depend on human raters for every test run. Google's pipeline grades thousands of multimodal responses against policy rules at once.
US 2026/0154578 A1
A feedback loop that catches hallucinations mid-generation lets the model revise its own output before output occurs, shifting from detection-after-the-fact to real-time correction within a single AI's reasoning chain.
US 2026/0148126 A1
Agents detect when peers lack needed context and retrain themselves mid-workflow, removing the need for human intervention after each failure.
US 2026/0148045 A1
The routing problem shifts from upfront model selection to mid-stream handoffs: detecting when a partial response signals the need for a different model before generation completes.
US 2026/0148012 A1
When multiple agents generate conflicting instructions for the same model, Google's system arbitrates by synthesizing them into a single unified prompt rather than letting contradictions reach the model.
US 2026/0147607 A1
Multiplying specialist models becomes practical when they stop each demanding their own copy of the base layer. IBM's approach pools that shared foundation, cutting the memory cost of running a whole team of fine-tuned variants.
US 2026/0141212 A1
Tool disambiguation in multi-agent systems: IBM's approach auto-detects when similar tools confuse routing decisions and reorganizes tool metadata to prevent mismatches, sidestepping manual descriptor rewrites.
US 2026/0141274 A1
Sequencing specialist models in order rather than running them in parallel sidesteps conflicts by letting each model work from the previous model's output.
US 2026/0141275 A1
Pooling shared neural network layers across multiple expert models cuts the memory footprint, making simultaneous specialist routing practical on single-GPU systems instead of requiring distributed clusters.
These filings focus on getting several AI models to work as a team instead of relying on one model for everything. That means routing questions to the right specialist, merging instructions when models disagree, and catching an AI's mistakes before a user sees them. Like all patents, they describe research directions, not confirmed products.
The watchlist currently includes filings from Google, IBM, and Microsoft, each approaching the same problem from a different angle. IBM's filings tend to focus on routing and shared infrastructure for running multiple models efficiently, while Google's cover routing, merging conflicting outputs, and testing AI behavior. Microsoft's filing looks at diagnosing failures inside chains of AI agents.
No. A patent filing shows a company has worked out a technical approach and wants to protect it, not that the feature is built, tested, or scheduled for release. Some ideas here, like routing tasks to specialist models or catching false answers automatically, may show up in products later, but the filings alone don't confirm timing.
When several companies file on similar problems around the same time, it usually means the industry has hit a shared bottleneck. Here, running one enormous model for every task is expensive and error-prone, so IBM, Google, and Microsoft are each patenting ways to split work across smaller, specialized models and catch mistakes along the way.
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →