Granted Patent B1

US 12,487,857 · App. 19/219,859 · Granted Dec 2, 2025

Inventors: Tilman Drerup (Berkeley, CA); Haixun Wang (Bellevue, WA); Sharath Rao Karikurve (Berkeley, CA)

Assignee: Maplebear Inc.

G06F9/5027

Abstract

Claims (69)

1 . A computer-implemented method, comprising:

creating an instance of a system artificial intelligence (AI) agent for an online system, wherein the system AI agent is configured to access a machine-learning language model;

detecting an instance of a user AI agent representing a user of the online system;

for one or more iterations:

receiving a message from the user AI agent,

providing one or more prompts for input to the machine-learning language model to request actions to execute for a current iteration based on the received message from the user AI agent,

parsing responses from the machine-learning language model to extract a set of selected actions and action inputs for the set of selected actions,

triggering, via an agent executor instance that is a compute process, execution of a set of respective tools corresponding to the selected actions with the action inputs,

generating a message for the user AI agent for the current iteration based at least on results of executing the set of respective tools, and

providing the generated message for the current iteration to the user AI agent;

extracting, from an interaction of the messages between the system AI agent and the user AI agent, a proposed agreement between the user and the online system;

performing one or more actions to execute the proposed agreement;

obtaining one or more records of previous or simulated interactions, wherein a record includes messages exchanged for a respective interaction between the system AI agent and another instance of a user AI agent, and tools executed for the interaction; and

training parameters of the machine-learning language model based on the one or more records of the previous or simulated interactions.

2 . The computer-implemented method of claim 1 , further comprising:

configuring one or more tools on an interface system, wherein the one or more tools include one or a combination of:

a first tool configured to access resources via an application programming interface (API), and

a second tool exposing functionalities of one or more task-based machine-learning models.

3 . The computer-implemented method of claim 1 , wherein training the parameters of the machine-learning language model further comprises:

computing a loss function including one or a combination of a first loss depending on a quality of user experience or a second loss depending on expected profits; and

backpropagating terms obtained from the loss function to update the parameters of the machine-learning language model.

4 . The computer-implemented method of claim 1 , further comprising:

providing an offer to purchase or one or more items through the online system to the user AI agent, wherein the message from the user AI agent for at least one iteration in the one or more iterations is a counter-offer to an offer message, and wherein the message generated for the user AI agent indicates a decision whether to accept the counter-offer or reject the counter-offer.

5 . The computer-implemented method of claim 1 , further comprising:

assigning a performance evaluation to each record, wherein the performance evaluation for each record indicates a degree of performance obtained by the system AI agent with respect to one or more criteria.

6 . The computer-implemented method of claim 5 , further comprising: providing the one or more records and performance evaluations for the one or more records in the prompts to the machine-learning language model for at least one iteration in the one or more iterations.

7 . The computer-implemented method of claim 1 , wherein the set of respective tools includes at least one of: an item description application programming interface (API) for retrieving details of an item, a delivery status API for retrieving details of a delivery status of an order, a machine-learning model for detecting fraud, or a machine-learning model for computing a likelihood a respective user will purchase an item.

8 . The computer-implemented method of claim 1 , wherein the one or more actions is one or a combination of invoking an application programming interface (API) to retrieve or change a compute resource, triggering a search query, or executing one or more machine-learning models.

9 . A non-transitory computer readable storage medium storing instructions that when executed by one or more computer processors cause the one or more computer processors to perform steps comprising:

creating an instance of a system artificial intelligence (AI) agent for an online system, wherein the system AI agent is configured to access a machine-learning language model;

detecting an instance of a user AI agent representing a user of the online system; for one or more iterations:

receiving a message from the user AI agent,

providing one or more prompts for input to the machine-learning language model to request actions to execute for a current iteration based on the received message from the user AI agent,

parsing responses from the machine-learning language model to extract a set of selected actions and action inputs for the set of selected actions,

triggering, via an agent executor instance that is a compute process, execution of a set of respective tools corresponding to the selected actions with the action inputs,

generating a message for the user AI agent for the current iteration based at least on results of executing the set of respective tools, and providing the generated message for the current iteration to the user AI agent;

extracting, from an interaction of the messages between the system AI agent and the user AI agent, a proposed agreement between the user and the online system;

performing one or more actions to execute the proposed agreement; and

obtaining one or more records of previous or simulated interactions, wherein a record includes messages exchanged for a respective interaction between the system AI agent and another instance of a user AI agent and tools executed for the interaction; and

training parameters of the machine-learning language model based on the one or more records of the previous or simulated interactions.

10 . The non-transitory computer readable storage medium of claim 9 , the instructions further causing the one or more computer processors to perform steps comprising: configuring one or more tools on an interface system, wherein the one or more tools include one or a combination of: a first tool configured to access resources via an application programming interface (API), and a second tool exposing functionalities of one or more task-based machine-learning models.

11 . The non-transitory computer readable storage medium of claim 9 , wherein training the parameters of the machine-learning language model further comprises:

computing a loss function including one or a combination of a first loss depending on a quality of user experience or a second loss depending on expected profits; and

backpropagating terms obtained from the loss function to update the parameters of the machine-learning language model.

12 . The non-transitory computer readable storage medium of claim 9 , the instructions further causing the one or more computer processors to perform steps comprising:

providing an offer to purchase or one or more items through the online system to the user AI agent,

wherein the message from the user AI agent for at least one iteration in the one or more iterations is a counter-offer to an offer message, and wherein the message generated for the user AI agent indicates a decision whether to accept the counter-offer or reject the counter-offer.

13 . The non-transitory computer readable storage medium of claim 9 , the instructions further causing the one or more computer processors to perform steps comprising:

assigning a performance evaluation to each record, wherein the performance evaluation for each record indicates a degree of performance obtained by the system AI agent with respect to one or more criteria.

14 . The non-transitory computer readable storage medium of claim 13 , the instructions further causing the one or more computer processors to perform steps comprising: providing the one or more records and performance evaluations for the one or more records in the prompts to the machine-learning language model for at least one iteration in the one or more iterations.

15 . The non-transitory computer readable storage medium of claim 9 , wherein the set of respective tools includes at least one of an item description application programming interface (API) for retrieving details of an item, a delivery status API for retrieving details of a delivery status of an order, a machine-learning model for detecting fraud, or a machine-learning model for computing a likelihood a respective user will purchase an item.

16 . The non-transitory computer readable storage medium of claim 9 , wherein the one or more actions is one or a combination of invoking an application programming interface (API) to retrieve or change a compute resource, triggering a search query, or executing one or more machine-learning models.

17 . A computer system, comprising:

one or more computer processors; and

a non-transitory computer readable storage medium storing instructions that when executed by the one or more computer processors cause the one or more computer processors to perform steps comprising:

creating an instance of a system artificial intelligence (AI) agent for an online system, wherein the system AI agent is configured to access a machine-learning language model;

detecting an instance of a user AI agent representing a user of the online system;

for one or more iterations:

receiving a message from the user AI agent,

providing one or more prompts for input to the machine-learning language model to request actions to execute for a current iteration based on the received message from the user AI agent,

parsing responses from the machine-learning language model to extract a set of selected actions and action inputs for the set of selected actions, triggering, via an agent executor instance that is a compute process,

execution of a set of respective tools corresponding to the selected actions with the action inputs,

generating a message for the user AI agent for the current iteration based at least on results of executing the set of respective tools, and

providing the generated message for the current iteration to the user AI agent;

extracting, from an interaction of the messages between the system AI agent and the user AI agent, a proposed agreement between the user and the online system;

performing one or more actions to execute the proposed agreement;

obtaining one or more records of previous or simulated interactions, wherein a record includes messages exchanged for a respective interaction between the system AI agent and another instance of a user AI agent and tools executed for the interaction; and

training parameters of the machine-learning language model based on the one or more records of the previous or simulated interactions.

18 . The computer system of claim 17 , the instructions further causing the one or more computer processors to perform steps comprising: configuring one or more tools and on an interface system, wherein the one or more tools include one or a combination of: a first tool configured to access resources via an application programming interface (API), and a second tool exposing functionalities of one or more task-based machine-learning models.

Assignments (2)

Continuity (1)

References Cited (23)