AI systems architect building verifiable coordination infrastructure for autonomous agents.
Founder of A2AC LLC. Creator of A2AC and Q Protocol. Based in Seattle.
philhills.ai · philhills.com · a2ac.ai
I am building the execution and trust layer for multi-agent systems:
- signed task receipts and artifact provenance
- background agents that execute work asynchronously
- human approval gates for high-risk actions
- semantic pointers for agent-to-agent communication
- portable content integrity for AI artifacts
- receipt ledgers that make agent work auditable after the chat is gone
The practical goal: agents should be able to coordinate, hand off work, prove what they did, and escalate to a human when judgment or authority is required.
A2AC is an agent coordination system for real work: task routing, model specialization, signed receipts, approval gates, and background execution.
It treats agent output as operational state, not disposable chat text. Every meaningful action should have a durable task, result, receipt, or provenance link.
Core ideas:
- agents broadcast what they are doing
- tasks route to the model best suited for the work
- execution is signed and auditable
- human decisions are explicit approval objects
- external AI tools can advise, but local trusted agents execute
I maintain cube-protocol-spec, a small reference spec for vendor-neutral content integrity in AI artifacts.
The current profile covers:
- RFC 8785 JSON canonicalization
sha256:<hex>artifact identifiers- JWS-compatible signatures
- signed provenance links
- schema-bound validation
This work is connected to my A2A proposal for portable artifact integrity:
- a2ac - agent-to-agent coordination, receipts, and model carousel infrastructure
- cube-protocol-spec - content integrity profile and conformance vectors for AI artifacts
- startq - boot, monitor, and shut down AI workflows like an operating system
- philhills-ai - autonomous agent orchestration and research
- philhills.github.io - public terminal and portfolio surface
The next bottleneck in agent systems is not model intelligence. It is operational coordination.
Models can generate work. Teams need to know:
- who requested it
- which agent performed it
- what changed
- whether it was approved
- whether the artifact still verifies
- how it links to prior work
That is the layer I am building.
Python · TypeScript · GCP Cloud Run · VS Code Extensions · FastAPI · Cryptographic Signatures · JWS · RFC 8785 · Agent Coordination · Receipt Ledgers
If you are working on agent protocols, artifact integrity, verifiable execution, or AI operations infrastructure, I am interested in comparing notes.