The same primitive — an accountable, anchored, independently verifiable agent identity — solves a different first problem for each institution that adopts it. Four concrete scenarios, each mapped to the mechanisms that make it provable rather than promised.
For payment operations, treasury, and compliance teams deploying agents into settlement, reconciliation, and regulated workflows.
A treasury agent moves funds between accounts and initiates vendor payments. The compliance desk cannot approve what it cannot attribute: which legal entity stands behind this agent, what mandate authorized this specific transfer, was the mandate live at the moment of action, and what tamper-evident record survives for the examiner two years from now?
PRAXIS answers each mechanically. The agent resolves to a depth-0 accountable root — a real legal principal, enforced at validation. Authority is a live, scoped, revocable mandate, default-deny at the point of action. Every state change emits an append-only event, so the full history is replayable from the chain itself, not reconstructed from logs the bank must be trusted to keep.
For platform, security, and governance teams running AI agents at scale across departments and vendors.
Agents proliferate faster than any access-review cadence. The organization needs one consistent answer to "does this agent hold authority for this action, right now?" — enforced identically whether the agent was issued by IT, a business unit, or an outside vendor, and revocable instantly when a mandate ends or a model is swapped.
Because identity is separated from the model, an agent can change runtimes without changing identity — the did:tz name, the delegation chain, and the anchored history all survive a model swap, while a provenance block records exactly which intelligence sat beneath the identity at each version. Chain-of-command is encoded, not documented.
For builders of agent-to-agent commerce, tokenized-asset platforms, and settlement networks.
An agent representing one firm transacts with an agent representing another. The counterparty needs to verify identity, authority, and standing without a shared intermediary and without taking any registry's word for it — the classic "trust the registry" single point of failure that reputation systems and directories cannot escape.
Verification here is stateless recomputation over public artifacts: re-hash the document against the on-chain commitment, walk the version chain to a constant genesis, check the signature, gate on policy. There is no score to farm and no directory to capture — capabilities are shared IRIs, so independently built indexes federate by merging quads, each side re-verifying for itself.
For standards authors and supervisors defining what agent identity and accountability must show.
Emerging guidance — NIST NCCoE, SP 800-63 assurance levels, the OWASP Non-Human Identity work — is converging on what institutions will be required to demonstrate for autonomous agents. The open question for any standard is whether a claimed control is actually checked, or merely cited.
PRAXIS treats a cited control without a wired check as a defect. Where it claims a standards property — DID-core shapes, credentials as hash references so inline PII is structurally impossible, validator-enforced assurance floors, PROV-O lineage — a validator or the conformance evaluator actually enforces it, and a published suite of golden byte-equality vectors lets any second implementation prove agreement in any language.
See the mechanisms in depth, or tell us which scenario is yours.