Genie Lite Architecture

Devpost Submission Draft — Genie Lite

Track

Professional Agents

Tagline

Human-led specialist routing with explicit authority, correction, recovery, and receipts.

One-line summary

Genie Lite is a Strands-based professional agent that carries human intent into specialist work without silently transferring human authority.

The problem

As professionals use AI for more than one-shot answers, the coordination burden grows fast. A person has to remember what they asked for, which capability did the work, whether a correction actually changed the work, what is safe to resume after interruption, and what the system has or has not actually proved.

Most agent experiences optimize for capability. Genie Lite focuses on inspectability under delegation.

Who it is for

Professionals, makers, creators, and small teams using AI across multi-step or specialist work where correction, handoff, recovery, and human authority matter.

What Genie Lite does

A human speaks to Echo, the conversational bridge. Echo carries the request to CREATE, one bounded specialist. CREATE returns specialist work through Echo. If the human corrects the direction, the next specialist brief and result must materially change. Receipts expose the route, return, authority state, and claim limits.

The contest seam is:

HUMAN → ECHO → CREATE → ECHO → HUMAN

Recovery restores useful state, but stale consequential authority does not silently resume.

Why it matters

The human should be able to carry less operational complexity without surrendering the decisions that matter.

Genie Lite makes four distinctions visible:

Conversation does not equal authority.
Routing does not equal execution.
Execution does not equal verification.
Recovery does not equal permission to resume.

Built with

Strands Agents
Amazon Bedrock AgentCore
Amazon Nova Pro (amazon.nova-pro-v1:0)
Python
AgentCore CodeZip deployment

Technical implementation

Genie Lite uses a Strands/AgentCore runtime with one visible specialist route. The implementation includes explicit state, authority, correction, recovery, and receipt behavior rather than burying those boundaries in the system prompt.

The project is deployed to Amazon Bedrock AgentCore in us-east-1; AgentCore reports the runtime as deployed and READY. A post-deploy diff reports no infrastructure drift.

What is proved

six deterministic credit-free tests pass
Echo→CREATE routing is visible
correction changes the next specialist brief/artifact
receipts expose route, return, authority state, and limits
stale consequential authority expires after interruption/recovery
AgentCore configuration validates
public repository is receiver-readable
AgentCore deployment completed successfully
runtime reports deployed / READY
post-deploy diff reports no differences

Current live-provider limitation

The first deployed Amazon Nova Pro invocation reached the runtime but returned a provider daily-token ThrottlingException.

We preserved that as an inspectable limitation instead of changing models, regions, or architecture to manufacture a green result. If the quota clears before submission close, the deployed normal/correction trace will be added as an evidence upgrade.

Until then, we do not claim a successful deployed model conversation.

What we do not claim

Genie Lite is not presented as production-ready, managed long-term memory, a generalized AI companion, or managed durable recovery. It does not automatically continue human authority after interruption.

What we learned

A useful human-led agent system needs more than capability. It needs boundaries that survive correction, handoff, and interruption.

The most important design lesson was:

The human can carry less complexity without giving the system more authority.

Repository

https://github.com/finious/genie-lite

Judge path

https://github.com/finious/genie-lite/blob/main/docs/JUDGE_START_HERE.md

Built With

agentcore-codezip
amazon-bedrock-agentcore
amazon-nova-pro
python
strands-agents

Genie Lite