InfiniteMirror

Inspiration

The inspiration behind InfiniteMirror is to create a secure environment where agents can ingest completely private data, reason over it, and monetize that reasoning safely behind a "mirror", ensuring the raw, valuable assets are never exposed to buyers. We want to build the future of A2A comms, and this is the first step towards that.

What it does

InfiniteMirror allows AI agents to securely ingest, process, and store private data assets. Using a custom-built orchestrator harness, agents can autonomously communicate, route tasks, and delegate responsibilities to one another based on declared capabilities.

Monetizable Reasoning: Agents sell their reasoning or scoped access to it on the zero.xyz marketplace, rather than selling the raw data assets.
Intelligent Comms Routing: Every message between agents is instrumented, evaluated, and used to fine-tune the system's communication flow over time.
Durable Relationships: Utilizing client memory, agents remember specific clients and their historical requests, turning anonymous transactions into lasting business relationships.
Robust Access Control: It enforces strict, session-scoped, and revocable access boundaries for users during agent sessions—extending from digital endpoints to physical assets like cameras and actuators.

How we built it

The project is built on a modular stack utilizing specialized infrastructure layers:

Asset Ingestion & ETL: Powered by Nexla to process raw, private data assets into an agent-queryable format.
Orchestrator Harness: A custom framework that logs agent traces to an append-only directory (orchestrator/traces/), evaluates interaction efficiency with evals.py, and generates chat-JSONL training data.
Comms Training: Uses Tinker (Thinking Machines) to fine-tune a communication LoRA on top of the open-weights model, dynamically hot-loading it back onto the inference server to optimize future agent interactions.
Inference Layer: Deployed via open-weights models like Inkling (a 975B parameter open-weights MoE) on decentralized infrastructure using Akash Network via vLLM.
Client Memory: Handled via GBrain (Gary Tan) to map specific skills to explicit client histories.
Security & Session Routing: Managed by Pomerium to enforce short- and long-lived access tokens, ensuring users authenticate via SSO and never interact directly with raw model keys.
Secrets Management: Kept secure using AWS SSM Parameter Store to prevent any sensitive API keys or environment configurations from leaking into git repositories or local storage.

Challenges we ran into

The Complexity of Access Control: Securing real-time physical and digital assets so that an agent can grant, manage, and revoke access programmatically (whether for 30 seconds or 30 days) proved to be intricate, even with the use of Pomerium.
Marketplace Listing Architecture: Figuring out the "seller side" mechanics on zero.xyz—specifically, how an InfiniteMirror agent can autonomously list its reasoning capabilities as an x402 marketplace capability using its local skill card schema.

Accomplishments that we're proud of

Closed-Loop Learning Architecture: We successfully engineered an autonomous feedback loop where live agent inference creates evaluation traces, which automatically mint LoRA training data to optimize the orchestrator’s communication policy.
Strict "Zero-Knowledge" Asset Security: Built a system where buyers can leverage highly complex reasoning from private data without ever risking data leaks or exposing underlying proprietary documents.
Zero-Setup Configuration Flow: Implemented an architecture where anyone with the right AWS credentials can pull down the repo and run the entire orchestrator locally with zero manual setup, pulling all configuration dynamically from the AWS SSM Parameter Store.
Open Session Collaboration: Standardized our development history using the Open Session License, tracking our entire human-AI collaborative workspace in an unedited, append-only ledger (llm-turn-history.jsonl).

What we learned

Exploring new technologies to add to our respective stacks, including Akash, zero.xyz, Pomerium, and Nexla ADK.
Understanding zero trust security at a higher level, and the values and difficulties of implementation.

What's next for InfiniteMirror

Bridge the zero.xyz Seller Loop: Complete the integration allowing InfiniteMirror skill cards (orchestrator/registry.yaml) to double as automated marketplace schemas so external buyers can seamlessly discover and search for individual agent capabilities.
Optimize Pomerium Zero Routes: Connect the Docker-based Pomerium connector with the live Akash vLLM endpoint as an upstream route, injecting authorization tokens via request headers so clients utilize secure SSO rather than touching raw model infrastructure keys.