Technology integrations
Connect the agents and infrastructure you already use.
Prudenze connects to agent frameworks, protocols, execution environments, workload identity, external policy and telemetry while preserving one governed decision path.
Implementation guidance and technical qualification
Prudenze connects evidence freshness, workload identity, policy and governed execution. Review the implementation guidance and recorded technical qualification for each integration boundary. Prudenze is a Hyperwise LLC product.
FreshCtx implementation guidance
FreshCtx 0.17.0 revalidates declared evidence at the action boundary. The pinned adapters below show where to enforce freshness in each framework. Their scope is freshness enforcement; business authorization and external outcome verification remain separate controls. Implementation links use source revision 8a7e7fb.
Pinned implementation, worked examples and verification source. The SDK ranges are declared compatibility ranges; exact installed versions and per-adapter qualification dates are not recorded in this inventory.
| Interface / boundary | Declared SDK range | Implementation evidence |
|---|---|---|
| Agno Tool hook | agno >=2.9,<3 | Adapter · Example · Verification |
| LangGraph Protected action node / checkpoint resume | langgraph >=1,<2 | Adapter · Example · Verification |
| OpenAI Agents SDK Custom function-tool input guardrail | openai-agents >=0.22,<1 | Adapter · Example · Verification |
| Google ADK before_tool_callback | google-adk >=1.26,<2 | Adapter · Example · Verification |
| ElevenLabs Python ClientTools handler | elevenlabs >=2.66,<3 | Adapter · Example · Verification |
| MCP Server tools/call extension | mcp >=2,<3 | Adapter · Example · Verification |
| A2A Receiving AgentExecutor | a2a-sdk >=1,<2 | Adapter · Example · Verification |
The framework conformance guide specifies current, stale, unavailable and unrelated-change cases. Google ADK's MCP 1.x dependency and MCP Guard's MCP 2.x boundary require separate environments. TheFreshCtx integration-author contract freshctx.pre_action.experimental.v1 is experimental and outside the stable FreshCtx API. A2A documentation describes receipt integrity and replay controls; business authorization is evaluated separately.
The September 27, 2026 qualification record covers FreshCtx 0.17.0 qualification. It is not an exact version matrix for every adapter. The named-host MCP report is specifically FreshCtx 0.9.0 with Codex CLI 0.151.0-alpha.7.2 on September 1, 2026; it does not establish current host compatibility.
Recorded ecosystem qualification
The technical qualification recorded September 29, 2026 records 24 passing verification checks across four interfaces. The supporting implementation and verification artifacts are available through controlled review. These results describe the qualified interfaces and versions below; production deployment requires environment-specific validation. Request the qualification material for review.
- OpenTelemetry — JS SDK 2.11.0; semantic-conventions 1.43.0; GenAI snapshot bcc7f9c2856fa7f4feb753f54d4ebba9455cc3dc
- Five recorded qualification checks: correlated spans, attributes, error paths and tenant/application correlation. Correlates supplied receipt identifiers.
- SPIFFE/SPIRE — SPIRE 1.15.3
- Six recorded qualification checks: X.509-SVID chain, trust domain, expiry and identity/authority separation. X.509-SVID qualification uses the isolated prudenze.test trust domain.
- Cedar — SDK/CLI 4.13.0; language 4.5
- Seven recorded qualification checks: permit/deny, malformed or unavailable input and Guardian precedence.
- OPA — OPA 1.21.1; Rego v1; OPA WASM SDK 1.10.0
- Six recorded qualification checks: compiled Rego, fail-closed handling and Guardian precedence.
For public runtime-action evidence, the existing OpenShell page documents seven verified control scenarios with pinned versions, reproduction steps and observed match/mismatch results.
Read the evidence at the right boundary
- Authorization record: what the controls permitted, escalated or denied, based on the supplied authority, policy and evidence.
- Observed execution: whether the protected handler or local mock action API was reached and what it recorded. A correlated receipt identifier alone is not this observation.
- Independently verified external outcome: reconciliation against an authoritative external system after the effect.
Technical scope: Adapter qualification covers the named interfaces and recorded versions. A decision or correlated receipt identifier does not establish an external effect. OpenShell results use a local mock payment API: no real funds moved or client systems were connected. Its comparator detects outcome differences after execution; it does not prevent, reverse or settle a payment.
For deployment planning, confirm the exact runtime versions, enforcement location, trust and credential lifecycle, retry/idempotency controls, authoritative outcome source and operational responsibilities. Commercial terms and vendor relationships are outside this technical qualification.
Agent Frameworks
05
Agno
Connect Agno agent and workflow actions to Prudenze governance boundaries.
LangGraph
Evaluate authority, evidence and policy before graph actions execute.
OpenAI Agents SDK
Apply Prudenze controls at agent handoffs and consequential tool calls.
Google ADK
Connect ADK agents and tools to governed decision and execution flows.
ElevenLabs
Govern consequential actions initiated through conversational voice agents.

