🏡 Austral Agent Stack

A multi-agent AI system for cost optimisation in Austral, NSW — managing a 9 kWh SolarEdge home battery and a Yamaha MT-10 motorcycle. Powered by Claude Sonnet, with a live Streamlit dashboard deployed to AWS.

Live dashboard → https://d31k8f0371hj3b.cloudfront.net

What it does

Every time you hit ▶ Run Analysis, 11 agents run — most concurrently — and produce a single plain-English recommendation:

☀️ How much solar will the 9 kW system generate today and tomorrow?
🔋 Should the battery export to the grid, soak solar, or preserve charge?
⛽ Is the cheapest P98 pump within 25 km worth riding to, after accounting for the fuel you'll burn getting there?
🌐 Is Brent Crude rising or falling — fill up now or wait?
🏍️ What's the best 2-hour window to ride the MT-10 today?
🚀 What phase is the moon in, and is the ISS passing overhead tonight?
🧠 Claude Sonnet synthesises everything into a 3-sentence briefing.

Architecture

Browser (HTTPS)
      │
      ▼
CloudFront (d31k8f0371hj3b.cloudfront.net)
      │  redirect-to-https, WebSocket pass-through
      ▼
Application Load Balancer  (port 80)
      │
      ▼
ECS Fargate  (512 vCPU / 1 GB, Streamlit port 8501)
      │  dashboard.py
      ▼
┌─────────────────────────── Orchestrator ──────────────────────────────┐
│                                                                        │
│  Concurrent ─────────────────────────────────────────────────────     │
│  ☀ SolarAnalyst  ⛽ FuelScout  🌐 MacroGeopolitics  🌤 RideScout  🚀 SpaceWatch  │
│                                                                        │
│  Sequential ─────────────────────────────────────────────────────     │
│  SolarAnalyst → BatteryManager → GridArbitrage                        │
│  FuelScout    → Logistics      → MT10Calculator                       │
│  All agents   ──────────────→  ClaudeAdvisor (LLM synthesis)         │
└────────────────────────────────────────────────────────────────────────┘

AWS Bedrock AgentCore (backend API, arm64 container)
  └── Same orchestrator, invoked via CLI or SDK
      └── Logs → CloudWatch /aws/bedrock-agentcore/AustralAgentCore

Agents

Agent	Role	API	Key needed
SolarAnalyst	24 h solar irradiance + kWh yield forecast	Open-Meteo	No
BatteryManager	GRID_EXPORT / SOLAR_SOAK / PRESERVE strategy	— (logic)	No
GridArbitrage	NSW NEM spot price → EXPORT / STORE / CONSUME	AEMO public	No
FuelScout	Cheapest P98 within 25 km, OAuth2 to NSW FuelCheck	NSW FuelCheck	Optional
Logistics	Riding distance from home to pump	OSRM	No
MT10Calculator	Is the detour profitable after the fuel cost?	— (math)	No
MacroGeopolitics	Live Brent Crude + AUD/USD trend sentiment	Yahoo Finance	No
RideScout	Hourly ride score (0–100) + best window	Open-Meteo	No
SpaceWatch	ISS position + moon phase + stargazing score	wheretheiss.at	No
ClaudeAdvisor	LLM synthesis of all agent outputs	Anthropic Claude	Yes

All agents have a synthetic fallback — the dashboard always loads even with no API keys configured.

Local Setup

1. Clone

git clone https://github.com/binzidd/agenttoagent.git
cd agenttoagent

2. Install dependencies

cd backend
pip install -r requirements.txt

3. Configure

cp backend/.env.example backend/.env

Edit backend/.env:

# Required for chat + Claude synthesis
ANTHROPIC_API_KEY=sk-ant-...

# Optional — falls back to synthetic data when absent
NSW_FUELCHECK_API_KEY=
NSW_FUELCHECK_API_SECRET=

4. Run

cd backend
streamlit run dashboard.py

Open http://localhost:8501, click ▶ Run Analysis.

Environment Variables

Variable	Default	Description
`ANTHROPIC_API_KEY`	—	Required. Claude Sonnet API key. Get one here
`NSW_FUELCHECK_API_KEY`	—	NSW FuelCheck key (OAuth2 client ID). Register free
`NSW_FUELCHECK_API_SECRET`	—	NSW FuelCheck OAuth2 client secret
`HOME_LAT`	`-33.93`	Home latitude (Austral NSW)
`HOME_LON`	`150.82`	Home longitude
`HOME_POSTCODE`	`2179`	Fuel search postcode
`SOLAR_SYSTEM_KW`	`9.0`	Installed solar capacity (kW)
`BATTERY_CAPACITY_KWH`	`9.0`	Battery capacity (kWh)
`FEED_IN_TARIFF_CENTS`	`5.0`	Solar export rate (c/kWh)
`BIKE_CONSUMPTION_L_100KM`	`7.5`	MT-10 fuel consumption
`BIKE_TANK_FILL_LITRES`	`15.0`	Typical fill volume

AWS Deployment — Full Walkthrough

This section documents every step taken to deploy this project to AWS, including every gotcha encountered along the way.

Overview of what gets deployed

Resource	Purpose	Cost estimate
ECR — `austral_agentcore`	arm64 container image for the backend	~$0.10/GB/month
ECR — `austral_dashboard`	amd64 container image for Streamlit	~$0.10/GB/month
Bedrock AgentCore runtime	Managed backend agent endpoint	Pay-per-invocation
ECS Fargate cluster + service	Runs the Streamlit dashboard	~$18/month (512 vCPU/1 GB, 24/7)
Application Load Balancer	Stable entry point in front of ECS	~$6/month fixed
CloudFront distribution	HTTPS termination, global CDN	~$0/month at low traffic
CloudWatch log group	14-day trace retention for AgentCore	Negligible

Prerequisites

AWS CLI v2 — install guide
Docker Desktop — running and logged in
jq — brew install jq
AWS credentials — configured via aws configure (or environment variables)

aws sts get-caller-identity   # confirm you're authenticated
docker info                   # confirm Docker is running

Step 1 — Create the AgentCore IAM execution role

AgentCore needs an IAM role to pull your container from ECR and write logs.

# Create the role with AgentCore as the trusted service
aws iam create-role \
  --role-name AustralAgentCoreRole \
  --assume-role-policy-document '{
    "Version":"2012-10-17",
    "Statement":[{
      "Effect":"Allow",
      "Principal":{"Service":"bedrock-agentcore.amazonaws.com"},
      "Action":"sts:AssumeRole"
    }]
  }'

# Attach the managed policy (gives AgentCore full access to itself)
aws iam attach-role-policy \
  --role-name AustralAgentCoreRole \
  --policy-arn arn:aws:iam::aws:policy/BedrockAgentCoreFullAccess

# Allow AgentCore to pull your ECR image
aws iam attach-role-policy \
  --role-name AustralAgentCoreRole \
  --policy-arn arn:aws:iam::aws:policy/AmazonEC2ContainerRegistryReadOnly

# Allow the container to write logs to CloudWatch
aws iam put-role-policy \
  --role-name AustralAgentCoreRole \
  --policy-name CloudWatchLogs \
  --policy-document '{
    "Version":"2012-10-17",
    "Statement":[{
      "Effect":"Allow",
      "Action":["logs:CreateLogGroup","logs:CreateLogStream","logs:PutLogEvents","logs:DescribeLogStreams"],
      "Resource":"arn:aws:logs:*:*:log-group:/aws/bedrock-agentcore/*"
    }]
  }'

Step 2 — Create the CloudWatch log group

aws logs create-log-group \
  --log-group-name /aws/bedrock-agentcore/AustralAgentCore \
  --region ap-southeast-2

aws logs put-retention-policy \
  --log-group-name /aws/bedrock-agentcore/AustralAgentCore \
  --retention-in-days 14 \
  --region ap-southeast-2

Step 3 — Deploy

The deploy.sh script handles everything from here.

# Deploy just the backend to AgentCore:
./deploy.sh agentcore

# Deploy just the Streamlit dashboard to ECS + ALB + CloudFront:
./deploy.sh dashboard

# Deploy both:
./deploy.sh all

On first run deploy.sh will:

Create ECR repositories with scan-on-push and a 3-image lifecycle policy
Build the Docker image and push to ECR
Create the AgentCore runtime (backend) or ECS service + ALB (dashboard)

Note: deploy.sh has the AWS account ID and default VPC/subnet IDs hardcoded for this project's environment. Fork users should update those values at the top of the file.

Gotchas encountered (and fixed)

These are real issues hit during deployment, documented so you don't repeat them.

❶ AgentCore only accepts `arm64` containers

ValidationException: Architecture incompatible for uri '...'.
Supported architectures: [arm64]

The Dockerfile builds linux/arm64. ECS Fargate (dashboard) stays on linux/amd64. They use separate ECR repos.

# AgentCore — arm64
docker build --platform linux/arm64 -f Dockerfile ...

# ECS dashboard — amd64
docker build --platform linux/amd64 -f Dockerfile.streamlit ...

❷ AgentCore control plane is `bedrock-agentcore-control`, not `bedrock-agentcore`

aws bedrock-agentcore is the data plane (invoke, sessions, memory). aws bedrock-agentcore-control is the control plane (create/update/list runtimes).

# ✅ Correct
aws bedrock-agentcore-control create-agent-runtime ...
aws bedrock-agentcore-control list-agent-runtimes ...

# ❌ Wrong — these commands don't exist here
aws bedrock-agentcore create-agent-runtime ...

❸ AgentCore runtime names cannot contain hyphens

Pattern constraint: [a-zA-Z][a-zA-Z0-9_]{0,47} — underscores only, no hyphens.

# ✅ Valid
--agent-runtime-name "AustralAgentCore"

# ❌ Invalid — causes ParamValidation
--agent-runtime-name "austral-agentcore"

❹ Environment variable JSON must be built with `jq`

If your .env file has quoted values (ANTHROPIC_API_KEY="sk-ant-..."), inline shell string interpolation into JSON produces double-quoted values and a parse error:

Invalid JSON: ... "ANTHROPIC_API_KEY": ""sk-ant-..."" ...

Fix: strip quotes from .env values, then use jq --arg to safely build the JSON object:

env_val() {
  local raw
  raw=$(grep "^${1}=" backend/.env 2>/dev/null | cut -d= -f2-)
  raw="${raw%\"}" ; raw="${raw#\"}"
  echo "${raw}"
}

build_env_json() {
  jq -n \
    --arg anthropic   "$(env_val ANTHROPIC_API_KEY)" \
    --arg fuel_key    "$(env_val NSW_FUELCHECK_API_KEY)" \
    --arg fuel_secret "$(env_val NSW_FUELCHECK_API_SECRET)" \
    '{ANTHROPIC_API_KEY:$anthropic, NSW_FUELCHECK_API_KEY:$fuel_key, NSW_FUELCHECK_API_SECRET:$fuel_secret}'
}

❺ ECS task execution role needs `logs:CreateLogGroup` separately

AmazonECSTaskExecutionRolePolicy does not include logs:CreateLogGroup. Tasks fail to start with:

ResourceInitializationError: failed to validate logger args: ...
AccessDeniedException: not authorized to perform: logs:CreateLogGroup

Fix: create the log group manually first, and add the permission:

aws logs create-log-group --log-group-name /ecs/austral-dashboard --region ap-southeast-2

aws iam put-role-policy \
  --role-name AustralDashboardTaskExecRole \
  --policy-name CloudWatchLogsCreateGroup \
  --policy-document '{
    "Version":"2012-10-17",
    "Statement":[{
      "Effect":"Allow",
      "Action":["logs:CreateLogGroup"],
      "Resource":"arn:aws:logs:ap-southeast-2:<account>:log-group:*"
    }]
  }'

❻ AWS App Runner requires a subscription (root credentials can't use it)

SubscriptionRequiredException: The AWS Access Key Id needs a subscription for the service

Solution: use ECS Fargate instead of App Runner. Same result, no subscription required.

❼ Security group descriptions must be plain ASCII

InvalidParameterValue: Value for parameter GroupDescription is invalid.
Character sets beyond ASCII are not supported.

Any em dash (–), smart quote, or non-ASCII character in the description causes this. Use plain ASCII only.

❽ ALB idle timeout must be raised for Streamlit WebSocket sessions

The default ALB idle timeout is 60 seconds. Streamlit uses persistent WebSocket connections. Sessions drop silently after 60 seconds of no interaction.

aws elbv2 modify-load-balancer-attributes \
  --load-balancer-arn <arn> \
  --attributes Key=idle_timeout.timeout_seconds,Value=3600

❾ ECS service cannot have a load balancer added after creation

You cannot call aws ecs update-service to add an ALB to an existing service. The service must be deleted and recreated with --load-balancers specified at creation time.

# Scale to 0, then delete
aws ecs update-service --cluster <cluster> --service <name> --desired-count 0
aws ecs delete-service --cluster <cluster> --service <name>
# Wait for DRAINING to complete (~30 s), then recreate
aws ecs create-service ... --load-balancers "..."

Invoking the AgentCore backend

# Full analysis
aws bedrock-agentcore invoke-agent-runtime \
  --region ap-southeast-2 \
  --agent-runtime-id AustralAgentCore-mhgPlO9T4b \
  --payload '{"action":"full_analysis"}' \
  response.json && cat response.json

# Natural language
aws bedrock-agentcore invoke-agent-runtime \
  --region ap-southeast-2 \
  --agent-runtime-id AustralAgentCore-mhgPlO9T4b \
  --payload '{"prompt":"Should I fill up today?"}' \
  response.json && cat response.json

Available actions: full_analysis · solar · fuel · ride · grid · chat

Observability — Tracing in CloudWatch

Every agent lifecycle event is logged as a structured JSON line to stdout, captured by the AgentCore runtime and forwarded to:

CloudWatch → Log groups → /aws/bedrock-agentcore/AustralAgentCore

Example log line:

{
  "ts": "2026-03-25T05:30:12.441Z",
  "level": "INFO",
  "event": "agent_complete",
  "agent": "SolarAnalyst",
  "data": { "forecast_yield_kwh_today": 38.2, "peak_generation_hour": 12 }
}

Tail logs live:

aws logs tail /aws/bedrock-agentcore/AustralAgentCore \
  --region ap-southeast-2 \
  --follow \
  --format short

CloudWatch Logs Insights query — see all agent completions for a run:

fields ts, agent, event, data.forecast_yield_kwh_today, data.price, data.score
| filter event in ["agent_complete", "workflow_complete"]
| sort ts asc

Project Structure

agenttoagent/
├── Dockerfile                     # AgentCore backend — linux/arm64
├── Dockerfile.streamlit           # Dashboard — linux/amd64
├── deploy.sh                      # One-command deploy to AWS
├── README.md
│
└── backend/
    ├── dashboard.py               # Streamlit dashboard (primary UI)
    ├── main.py                    # AgentCore entry point + structured logging
    ├── orchestrator.py            # Concurrent + sequential agent coordination
    ├── api.py                     # FastAPI + WebSocket server (optional)
    ├── config.py                  # Pydantic settings
    ├── requirements.txt
    ├── .env.example
    │
    └── agents/
        ├── solar_analyst.py       # Open-Meteo irradiance forecast
        ├── battery_manager.py     # Charge/export/preserve strategy
        ├── grid_arbitrage.py      # AEMO NEM spot price arbitrage
        ├── fuel_scout.py          # NSW FuelCheck OAuth2 + Haversine filter
        ├── logistics.py           # OSRM open routing
        ├── mt10_calculator.py     # Detour profitability maths
        ├── macro_geopolitics.py   # Brent Crude + AUD/USD via yfinance
        ├── ride_scout.py          # Hourly weather ride scorer
        ├── space_watch.py         # ISS position + moon phase (wheretheiss.at)
        └── claude_advisor.py      # Claude Sonnet synthesis + streaming chat

Data Sources

Data	Provider	Free	Docs
Solar irradiance + weather	Open-Meteo	✅	open-meteo.com
Riding distance	OSRM	✅	project-osrm.org
Brent Crude + AUD/USD	Yahoo Finance via yfinance	✅	—
NEM electricity spot price	AEMO	✅	aemo.com.au
ISS position	wheretheiss.at	✅	wheretheiss.at
P98 fuel prices	NSW FuelCheck	✅ (key required)	api.nsw.gov.au
LLM synthesis + chat	Anthropic Claude	Paid	console.anthropic.com

Troubleshooting

`ModuleNotFoundError: No module named 'pydantic_settings'`

pip install pydantic-settings httpx yfinance anthropic plotly streamlit

Chat returns "⚠️ Set ANTHROPIC_API_KEY"

Add to backend/.env:

ANTHROPIC_API_KEY=sk-ant-...

Fuel prices show `(est.)` badge

No FuelCheck key configured — app falls back to synthetic prices. Register free.

AgentCore invocation returns immediately with no result

Check that the runtime status is READY:

aws bedrock-agentcore-control list-agent-runtimes --region ap-southeast-2

If status is UPDATING, wait ~2 minutes after a deploy before invoking.

Dashboard CloudFront URL returns 503

ECS task may have restarted (new public IP). The ALB health check re-registers it within ~30 seconds. Wait and refresh.

Name		Name	Last commit message	Last commit date
Latest commit History 18 Commits
backend		backend
.gitignore		.gitignore
Dockerfile		Dockerfile
Dockerfile.streamlit		Dockerfile.streamlit
README.md		README.md
deploy.sh		deploy.sh

Folders and files

Latest commit

History

Repository files navigation

🏡 Austral Agent Stack

Table of Contents

What it does

Architecture

Agents

Local Setup

1. Clone

2. Install dependencies

3. Configure

4. Run

Environment Variables

AWS Deployment — Full Walkthrough

Overview of what gets deployed

Prerequisites

Step 1 — Create the AgentCore IAM execution role

Step 2 — Create the CloudWatch log group

Step 3 — Deploy

Gotchas encountered (and fixed)

❶ AgentCore only accepts arm64 containers

❷ AgentCore control plane is bedrock-agentcore-control, not bedrock-agentcore

❸ AgentCore runtime names cannot contain hyphens

❹ Environment variable JSON must be built with jq

❺ ECS task execution role needs logs:CreateLogGroup separately

❻ AWS App Runner requires a subscription (root credentials can't use it)

❼ Security group descriptions must be plain ASCII

❽ ALB idle timeout must be raised for Streamlit WebSocket sessions

❾ ECS service cannot have a load balancer added after creation

Invoking the AgentCore backend

Observability — Tracing in CloudWatch

Project Structure

Data Sources

Troubleshooting

ModuleNotFoundError: No module named 'pydantic_settings'

Chat returns "⚠️ Set ANTHROPIC_API_KEY"

Fuel prices show (est.) badge

AgentCore invocation returns immediately with no result

Dashboard CloudFront URL returns 503

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

❶ AgentCore only accepts `arm64` containers

❷ AgentCore control plane is `bedrock-agentcore-control`, not `bedrock-agentcore`

❹ Environment variable JSON must be built with `jq`

❺ ECS task execution role needs `logs:CreateLogGroup` separately

`ModuleNotFoundError: No module named 'pydantic_settings'`

Fuel prices show `(est.)` badge

Packages