Anumati / ACAP
Consent and Adherence. Per-clause usage-policy evaluation with proof of adherence, not just proof of acceptance. Read more →
Agent Protocol Stack
Independent, composable protocol extensions for the A2A and MCP agent stack. Each addresses a structural gap that current protocols leave open.
Phala
Welfare Feedback. Principal-declared outcome quality routed back through the agent network. The principal defines what good means. Read more →
Pratyahara / NERVE
Behavioral Integrity. Real-time drift detection, collusion detection, and network-level health monitoring for compromised agents. Read more →
Sauvidya / PACE
Accessibility. Principal capability profiling, adapted interaction modalities, and consent capacity checks for vulnerable populations. Read more →
Pramana
Verification. Wrap every consequential agent output in a typed ClaimAttestation with a verify() operation against the recorded source. Standardizes the verification artifact across the patterns already in production (RAG, tool-augmented traces, generator-verifier loops) and replaces probabilistic verdicts with a deterministic contract where the claim type admits one. Read more →
Abhyasa
Reliable Delivery. Custody transfer of governance obligations under a deliver-or-report guarantee, so a consent decision or a corrective welfare signal is never silently lost over a lossy or partitioned channel. The delivery layer the other protocols ride on. Read more →
Samsara
Lifecycle and Memory. What an instance carries across its own death: not what it did, but what it became. Episodes are destroyed at shutdown by contract; only consent-gated, outcome-signed dispositions persist, and any life’s own contribution is exactly excisable by replaying the merge without it. Tested by a preregistered battery on two live models. Read more →
Transports
Section titled “Transports”These are governance extensions, not a transport of their own. They bind to the
agent stack’s transports: A2A (agent↔agent) via capabilities.extensions[]
and MCP (agent↔tools) via _meta. They also bind to
AG-UI, the agent↔human
transport — the edge where a person enters a run to grant consent, respond to a
capacity check, or rate an outcome.
Governance over AG-UI
Human-in-the-loop binding. How a governance decision projects onto AG-UI’s interrupt mechanism, with an interactive walk-through of one governed run. Read the binding →