GreenRoute Human Agent

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for GreenRoute Human Agent

Inspiration

AI workloads are often routed to the largest available model by habit. That can ignore data-region rules, required quality and the uncertainty behind energy estimates. GreenRoute Human Agent turns that choice into a reviewable decision instead of a hidden default.

What it does

A person describes a workload and sets quality, data region and expected output length. A Strands agent calls an authoritative deterministic tool that classifies the task, applies region policy before optimisation, enforces a quality floor, selects an adequate route and returns a prompt-free audit record. If the policy combination cannot be satisfied, the tool refuses rather than weakening a constraint silently. Every estimate carries an explicit uncertainty range and human-review checklist.

How we built it

The new agent layer uses the Strands Agents SDK with its official Ollama provider, so it can run locally without an AWS deployment, billing account or paid API. Python implements the deterministic routing engine, refusal path and unit tests. The local language model may explain and sequence the workflow, but it cannot override the tool output. A browser-based audit interface provides a transparent visual companion and JSON export.

Safety and human control

GreenRoute never describes its estimates as direct measurements or verified avoided emissions. It does not store the workload prompt in the audit record. Provider availability, quality and production coefficients stay in a human-review list until verified. Unknowns remain unknown.

Challenges

The main challenge was separating helpful agent behaviour from decisions that require deterministic enforcement. We solved this by making the policy tool authoritative and using the model only for orchestration and explanation. We also designed a zero-cost local execution path instead of requiring cloud billing.

Accomplishments

• Working deterministic routing and explicit refusal path
• 3 passing unit tests
• Strands tool integration with local Ollama
• Public MIT-licensed repository and reproducible setup
• Architecture diagram and honest reuse disclosure

What we learned

Human-centred agents are strongest when they expose limits and escalation points. A smaller, inspectable decision with a clear refusal can be more useful than a confident answer from a larger model.

What's next

Replace illustrative coefficients with licensed provider data, add evaluation datasets for quality thresholds, expose signed audit records and test the workflow with teams responsible for AI governance.

Reuse disclosure

The browser Audit Lab was first built during the DevNetwork API + Cloud + AI Hackathon 2026. The Strands/Ollama agent wrapper, Python policy engine, refusal path, tests and agent architecture were created during the Agents for Humans submission period. This entry is the new human-review agent workflow; the earlier interface is disclosed as a visual companion.

Built With

• css3
• html5
• javascript
• ollama
• python
• strands-agents