StillOn

• 

  Morning board

• 

  Interrupt scene

• 

  Architecture

Inspiration

Food banks already help people enroll in SNAP, Medicaid, and LIHEAP. The part that still breaks is quieter: six months later a recertification letter sits on the counter, the drop date passes, and the household is hungry, uninsured, or cold. That is procedural churn, not a discovery problem.

I did not want another chatbot that finds a pantry. Those exist, and this hackathon is full of them. I wanted the overnight shift that does not exist — a desk that works the caseload while the caseworker sleeps and only wakes a human when a real decision is required.

Harbor Light is a demo pantry in Columbus, Ohio. The twelve households are synthetic. The chore is real.

What it does

StillOn ranks every enrollment by drop date in Python. Quiet cases never enter the model. For households due in twelve days, or already flagged, a Strands agent loads the file, matches evidence to the notice, and either drafts a PDF packet or pauses.

In the morning the caseworker opens a board, not a prompt box. Needs you is two or three buttons. Ready to file is a silent PDF. A click resumes the same Strands session. Filing is an interrupt. The model cannot submit on its own.

How I built it

The product is a harness around a small model, not a prompt with a UI.

• Facts are code. Drop dates, urgency bands, and paystub freshness live in deadlines.py and matching.py. The LLM does not invent a due date.
• Judgment is the Strands agent. Six tools: household file, match, draft packet, escalate, submit, priority queue. Overnight it may draft. It may not file.
• Authority is a hook. CaseworkerGate listens for BeforeToolCallEvent on escalate_decision and submit_packet. The paused session is stored and resumed after a button click.
• The board is FastAPI + a static morning UI. No chat input.
• The model is swappable. Local NightDeskModel for a demo with no AWS. Amazon Nova Lite in Ohio (us-east-2) when STILLON_USE_BEDROCK=1. Same tools. Same interrupts.
• AgentCore wraps the same night / board / decide loop in deploy/agentcore_app.py.

Challenges I ran into

Blank Slate is not the classic console. The project region is Ohio, not N. Virginia, and chasing us-east-1 wasted time. Nova Lite in the playground was the first model that actually answered. Short-lived Bedrock API keys meant the night desk had to keep working when the key expired — that is why the local model provider exists.

Nova sometimes stuffed the choice list into a single control labeled options. A caseworker cannot click JSON. I stopped trusting the model for buttons. Python policy now owns the two choices (stale stub, missing signature, income change, already dropped). The model can still write the question.

One bad tool call used to kill the whole night. Each household is now isolated. A failure parks that file and the rest of the caseload still runs.

The harder design challenge was saying no to a prompt box. A chat demo is easier to record and easier to lose on. StillOn only works if the morning board stays quiet.

What I learned

Determinism is what this hackathon is scoring, not a longer feature list. Deadlines belong in Python. Interrupts belong in hooks. The model should pick the next tool, not invent policy.

I also learned that a neighbor agent is a caseload, not a personal assistant. If StillOn works, twelve families keep food, coverage, or heat they already earned. If it chats, they still drop.

What's next

The morning board is live. AgentCore Runtime is live in Ohio (arn:aws:bedrock-agentcore:us-east-2:750390206396:runtime/StillOn_StillOn-C6Gf1qBQlK). Next is a real pantry: real notices, same rule — the model drafts, the human files.s.

Built With

• agentcore
• amazon-bedrock
• amazon-nova
• amazon-web-services
• fastapi
• pydantic
• python
• reportlab
• strand
• strands-agents
• uvicorn