A multi-agent mathematical modeling system built on Amazon Bedrock AgentCore — give it an open-ended math modeling problem, and it automatically runs the full pipeline of Problem Analysis → Mathematical Modeling → Computational Solving → Solution Reporting, streaming the reasoning and results in real time.
English | 简体中文
MathModeler is an end-to-end multi-agent application that runs entirely on Amazon Bedrock AgentCore. It decomposes a complex modeling task into specialized roles, coordinated by an orchestrator agent that drives collaboration along the task's dependencies:
- Analyst — understands the problem, decomposes it into sub-problems, and builds the dependency DAG between them.
- Modeler — retrieves candidate methods from a hierarchical modeling-method library and iteratively refines the mathematical model in an actor-critic loop.
- Solver — automatically writes and executes code, retries on errors, and produces computational results.
- Reporter — aggregates every stage's output into a structured, complete solution report.
All collaboration, reasoning, and solving is hosted on AgentCore — no need to build your own container orchestration or session-state management.
The system makes full use of the core capabilities of Amazon Bedrock AgentCore:
| AgentCore capability | Role in this project |
|---|---|
| AgentCore Runtime | Each agent (the Orchestrator + 4 sub-agents) is deployed as an independent serverless Runtime, exposing /invocations (POST) and /ping (GET) via FastAPI, with SSE streaming, running on LINUX_ARM64. |
| AgentCore Code Interpreter | Provides a controlled sandbox execution environment for the Solver to safely run model-generated solver code. |
| AgentCore Memory | Manages short-term session events and long-term preferences, enabling context memory across steps. |
| Amazon S3 (document bus) | Agents exchange intermediate artifacts through S3 (organized by session_id) for decoupled collaboration. |
The web portal is served by a FastAPI container on AWS Fargate behind an Application Load Balancer (HTTP). The ALB streams chunked/SSE responses without buffering, so the four-stage progress reaches the browser in real time; the portal also proxies /api/solve to the Orchestrator Runtime (InvokeAgentRuntime, SigV4) and gates access with a single admin login.
Agents are organized with the Strands Agents framework in an agents-as-tools pattern. The Orchestrator both constructs the supervisor agent and drives the sub-agents through a deterministic pipeline.
Prerequisites
- An AWS account, with credentials configured locally — e.g.
aws configure(orAWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY/AWS_SESSION_TOKENenvironment variables, or an SSO profile). Without valid credentials the deployment will fail. - The target region (e.g.
us-west-2) must have model access enabled in Amazon Bedrock for the models this project uses. - Node.js 18+ and Docker installed (CDK builds the agent container images locally; an
arm64/Graviton build host is recommended since the runtimes targetLINUX_ARM64).
Deploy
# 1. Configure your AWS credentials first
aws configure # or export AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_SESSION_TOKEN
# 2. Deploy with CDK (supply the portal admin password via CDK context)
cd mathmodeler/infra
npm install
npx cdk bootstrap # first time only, per account/region
npx cdk deploy -c adminUser=admin -c adminPassword='<your-password>'After deployment, open the PortalEndpoint output (an http://<alb>.<region>.elb.amazonaws.com URL),
log in with the admin credentials you provided, and submit a problem to watch the four stages stream live.
Note: Without a custom domain / ACM certificate the ALB serves over HTTP (the default
*.elb.amazonaws.comname cannot be issued a certificate), so the login password is sent in clear text — acceptable for a demo. Add an HTTPS listener once you have a domain.
Thanks to MM-Agent for the methodology that inspired this project.