Agents

Agent systems—validated, certified, and releasable

Agent frameworks can plan, call tools, and execute workflows. In regulated environments that’s not enough. The missing ingredient is institutional authority control: who may ask, what evidence is required, which policy bundle governs, and who may release. PtolemyAI provides that layer with CRSP / NRC: validator-based AER loops plus on-chain system certification.

Agent System A — Anthropic Cowork (skill-based plugins)

Desktop-first agent runtime with domain “skills” and connectors. Strong for repeatable workflows (review, drafting, ops checklists) and fast adoption by teams.

Skill plugins Local execution Tool connectors Workflow automation
  • Agent can plan multi-step work and operate across files/tools
  • Skills define repeatable procedures (e.g., financial statements, legal review)
  • Great “Type-II capability surface” for enterprise teams
What CRSP adds: Ask/Evidence/Propose/Entitle/Release gating, defeaters, role attestations, and on-chain release proofs.

Agent System B — Gemini ADK (comparable enterprise agent runtime)

Enterprise agent development kit for building tool-using, orchestrated workflows (routing, actions, multi-step plans) inside a controlled application runtime.

ADK orchestration Tool calling Typed actions Deployment pipelines
  • Developer-driven agent workflows with structured tool interfaces
  • Strong integration posture for internal systems and controlled action spaces
  • Another “Type-II capability surface” that benefits from institutional governance
What CRSP adds: independent evaluator gates, policy-hash binding, entitlements, and audit-grade release discipline.

CRSP / NRC: the validator platform for agent systems

CRSP/NRC is not another agent framework. It is a comprehensive validator and certification layer that sits above agent runtimes. It ensures “agents” never become authorities: they propose—institutions release.

AER Validator Loop Deterministic Validators Defeater Engine Role Attestation On-chain System Certificate Safe Multisig

1) AER gating (validator-first)

Every workflow is decomposed into Ask → Evidence → Propose → Entitle → Release. Each phase produces hash-stable bundles for audit replay.

No silent outputs Proposals only Release requires entitlement

2) Independent evaluation of vendor agent behavior

CRSP can use a separate evaluator model (e.g., OpenAI) to score compliance of an Anthropic (or Gemini) agent run. Violations map to deterministic defeaters—no “repair loops” in system certification mode.

Cross-vendor evaluation Defeater determinism No propose repair

3) On-chain system certification (NRC)

Templates, policy bundles, and build hashes are anchored on-chain. Releases are cryptographic events: payload hashes + role attestation hashes + event chaining for end-to-end traceability.

Policy hash match Audit replay Institutional accountability
Result: You can adopt “agent systems” without inheriting “agentic liability.” The institution remains the authority.

Vendor layer vs institution layer

Agent platforms govern execution. CRSP/NRC governs authority.

  • Vendor runtime: planning, tool calls, workflow execution
  • CRSP: phase gates + evidence requirements + defeaters
  • NRC: system certificate + policy hash binding + release events
You keep your preferred vendor—CRSP/NRC makes its outputs releasable in regulated reality.

Palantir as operational control plane

Many enterprises already run governance workflows inside a platform control plane.

  • Use Palantir (AIP/Foundry) for permissions, lineage, ontology-bound actions
  • Wrap agent outputs with CRSP/NRC for entitlement and on-chain release certification
  • Operational governance + institutional authority = regulator-grade posture
Platform ops reduce drift; CRSP/NRC prevents authority leakage.

System certificate: what gets anchored

A system certificate binds what the agent is allowed to do to how outputs may be released. It is a contract for institutional accountability—not a model “trust badge”.

build_hash policy_bundle_hash tool allowlist data scope validator version release class
Certification makes workflows inspectable: you can replay what happened, prove what governed it, and show who authorized release.