FLEX Autopilot

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 FLEX Autopilot

Inspiration

Cloud training decisions are usually made from current utilization alone, while energy availability, weather, workload deadlines, and cost can all change before a workload starts. We built FLEX Autopilot to turn that changing evidence into a decision that can be replayed, challenged, and either admitted or rejected.

What it does

FLEX takes one objective and hard constraints, then runs a four-role workflow:

1. Sentinel validates the event and input sources.
2. Analyst retrieves point-in-time evidence and interprets operating conditions.
3. Optimizer calls FLEX Quant to compute the candidate schedule, region, energy, cost, and deadline tradeoffs.
4. Auditor independently checks the numeric result, source lineage, and grounding evidence before approving the proposal or returning HOLD.

FLEX OS then verifies the exact approved envelope, resource ceiling, and rollback contract. The live demonstration shows a private Google Compute Engine training canary through CREATE → TRAIN → VERIFY → DELETE, preserving a content-addressed receipt.

How we built it

The agent layer uses Google ADK SequentialAgent with Gemini 3.7 Flash on Vertex AI. The agents communicate through HYPHA, a typed envelope protocol carrying sender, receiver, input/output kind, authority, lineage, and hashes.

The evidence plane uses three Vertex AI Agent Search stores: External Intelligence, FLEX Evidence, and FLEX Operations. Results are re-ranked with the Vertex Ranking API, summarized with Vertex GenerateContent, and evaluated with Check Grounding. Generated language can explain evidence, but cannot supply a Quant number.

FLEX Quant is a deterministic Rust runtime that evaluates identical inputs with checked arithmetic and replayable receipts. FLEX OS performs separate admission checks. Cloud Run, Pub/Sub, Firestore, Cloud Storage, KMS, Artifact Registry, Cloud Build, and private Compute Engine provide the operating substrate.

The data foundation preserves NASA POWER, NOAA GFS/GEFS, and public KPX evidence as immutable Bronze data, deterministic Silver normalization, and leakage-safe Gold DecisionTuples across 25 operating regions.

Challenges we ran into

The core challenge was separating language reasoning from numerical authority. We deliberately rejected a design where retrieval or a language model could invent a cost, energy, or scheduling value. The system instead treats retrieval, reasoning, optimization, verification, and execution as separately authorized layers.

Accomplishments that we're proud of

• Four typed Gemini roles with machine-checkable handoffs.
• Three Agent Search stores with ranking, grounded generation, and claim-level verification.
• Deterministic Quant isolation: generated text is never a numeric input.
• 25 operating regions and 100 real historical Gold decision cases.
• 100/100 clean workflows and 100/100 injected fault cases handled correctly in the measured scorecard.
• Seven hard gates with zero unauthorized effects in the measured candidate set.
• A real private Google Cloud canary with a retained, content-addressed receipt.

What we learned

Agents become substantially more useful when they can prove why they stopped. Retrieval, model reasoning, numeric optimization, independent verification, and execution admission require different authorities. Google Cloud's AI and operating services let those boundaries become a working system, rather than a chatbot interface.

What's next

We are extending FLEX with customer-authorized scheduler adapters, multi-node validation, and additional telemetry sources.

Built With

• ai
• cloud
• compute
• google
• machine
• vertex