FINAL HACKATHON VERSION — August 31, 2026

InnerOS — The Self-Healing Agentic Operating System

This is the completed frozen submission for the All Things Agentic Hackathon 2026 — Fortified Enterprise Fleet.

Canonical repository: https://github.com/Rafa-Innerchispa/innerops-agentic-platform

Frozen hackathon branch: https://github.com/Rafa-Innerchispa/innerops-agentic-platform/tree/hackathon-freeze-20260831

Exact immutable final commit: https://github.com/Rafa-Innerchispa/innerops-agentic-platform/commit/c8e2f59abfcaa6171c1631d750f22a209fda61b5

Judges should evaluate the frozen branch or exact commit above. Normal product development may continue on other branches without changing the submitted version.

Live Judge Console: https://inneros.creatorcore.ai/app/judge

Judge login: https://inneros.creatorcore.ai/app/login?judge=1

Username: HACKATHON-JUDGE or DEVPOST-JUDGE Password: demo123

Demo video: https://youtu.be/IV12J4VcxEg

WHAT INNEROS DOES

InnerOS / ARIA is a local-first, self-healing enterprise agent operating layer for coordinating real work across persistent state, specialized agents, sovereign local AI, Gemini, Google Cloud, A2A, MCP and bounded operational tools.

Its operating loop is: signal → relevance → memory → decision → delegation → execution → verification → recovery → evidence → learning

InnerOS is designed to reduce the human coordination burden: agents can discover capabilities, route work, execute through scoped tools, preserve context, recover stalled execution and produce auditable evidence.

FORTIFIED ENTERPRISE FLEET

The final system demonstrates:

• canonical agent and capability discovery • A2A Agent Cards and multi-agent coordination • durable task lifecycle state • persistent operational context • correlation IDs, ownership, revisions and heartbeats • bounded local/cloud Resource Fabric routing • scoped identities, tools, secrets and approvals • recovery and verification loops • Global Live Trace observability • explicit PASS / PARTIAL / FAIL execution truth

GOOGLE REQUIREMENTS — COMPLETED

The final hackathon build uses and demonstrates:

• Gemini 3.5+ • Google Agent Development Kit (ADK) • Google GenAI SDK (google-genai) • Vertex AI • FunctionGemma / Gemma • Cloud Run • Firestore • Pub/Sub • Google Cloud identity / policy patterns • Model Armor / agentic-defense controls where integrated

Verified submission evidence includes Gemini execution, Google agent-framework integration, Cloud Run deployment, Firestore persistence, Pub/Sub events, Google Cloud observability and Vertex AI / FunctionGemma execution evidence.

FUNCTIONGEMMA / VERTEX AI

For the final submission, FunctionGemma Judge Test 3 was connected to the live Vertex AI route and successfully exercised after the final production hotfix.

The final hotfix removed the stale dedicated Vertex DNS dependency, uses current endpoint discovery and prevents an application HTTP 502 from incorrectly masking the real probe result.

The final live Judge integration hotfix is preserved at: Rafa-Innerchispa/innerspark-workforce-ai@6808383d9f098839e5754a8405e010dd9bd28601

SEVEN JUDGE PROOFS

The Judge Console exposes seven independently runnable proofs:

  1. system health and fresh correlation ID
  2. A2A discovery and connectivity
  3. live FunctionGemma / Vertex AI execution
  4. Gemini generation with a real downloadable PDF artifact
  5. ARIA natural-language interaction
  6. sovereign local AMD model inference
  7. bounded multi-agent dispatch with durable Global Live Trace evidence

ARIA and Global Live Trace are visible together so judges can inspect what ran, where it ran and what evidence was produced.

LOCAL-FIRST RESOURCE FABRIC

The sovereign AI plane includes AMD Radeon AI PRO R9700-class hardware, ROCm, vLLM and Qwen3-Coder-class local inference, plus lighter local runtimes and deterministic tools.

Agents request capabilities rather than hard-coding a machine or vendor. This allows the same bounded workflow to route across local infrastructure or Google Cloud according to capability, privacy, cost, load and risk.

SECURITY AND GOVERNANCE

InnerOS uses scoped and allowlisted tools, server-side secrets, tenant and identity boundaries, repository locks, isolated worktrees, approval gates for sensitive mutations, bounded retries and budgets, stalled-worker detection, verification gates and explicit execution truth.

REPRODUCIBILITY

The frozen repository includes:

• README.md — final architecture and setup • JUDGES_START_HERE.md — judge entrypoint, credentials and testing path • docs/ALL_THINGS_AGENTIC.md — Google requirements and final evidence mapping • architecture and security documentation • environment and dependency setup • API / MCP startup instructions • regression-test instructions

To reproduce the submitted source:

git clone https://github.com/Rafa-Innerchispa/innerops-agentic-platform.git cd innerops-agentic-platform git checkout hackathon-freeze-20260831

PRE-EXISTING WORK DISCLOSURE

The broader ecosystem already contained real products and infrastructure before the event. The hackathon contribution is the unified InnerOS / ARIA enterprise fleet layer and the substantial work completed during the submission period, including durable coordination, expanded A2A interoperability, Google-native execution, Resource Fabric routing, Judge Console / Global Live Trace, verification and recovery loops, persistent evidence and governance hardening.

FINAL PRINCIPLE

InnerOS notices what matters, chooses the right bounded intelligence and execution resource, does what it safely can, proves what happened and brings the human back only when human authority is actually needed.

Built With

  • a2a
  • cloud-run
  • firestore
  • functiongemma
  • gemini-3.5+
  • google-adk
  • google-genai-sdk
  • mcp
  • mongodb
  • next.js
  • pub/sub
  • qdrant
  • qwen3-coder
  • rocm
  • vertex-ai
  • vllm
Share this project:

Updates

Submission history