Skip to main content
Upgrade related Agentium packages together to 4.11.0. Existing Agent question and answer shapes remain supported. If you are upgrading from before 4.9, review the 4.9 compatibility guide as well.
Install only the packages your application uses. See installation for optional SDKs and supported Node versions.

​
Recheck Decisions classifications

Earlier Decisions adapters converted system, user, and assistant messages into user text with role labels. Version 4.11 preserves real system, developer, user, and assistant roles, plus paired function-call history. Existing Agent.run() and stream() calls keep their question and answer shapes, but stronger role semantics can change results. Run representative labeled examples through your existing thresholds before switching production traffic. Continue handling individual refusals explicitly. A successful request can still contain refused questions.

​
Choose native evidence when needed

Use new OpenAIDecisionsProvider(...).evaluate() for JSON evidence and historical custom-tool calls. The registry-backed openaiDecisions() factory keeps its generic ModelProvider type for Agent integrations. Questions accept string instructions or text/image parts, with at most 400 questions and 128 images across the complete request. Hosted HTTP(S) image URLs are forwarded to OpenAI; Agentium does not fetch them. Native tool history is evidence and does not execute tools. See Decisions for JSON limits and explicit prompt-cache controls. Direct evaluation returns usage but does not create an Agent ledger attempt. Use an Agent when you need its accounting lifecycle, or record direct usage in your application’s accounting integration.

​
Review cost estimates

Decisions now separates ordinary input, cache reads, and cache writes. Missing cache counters remain unknown instead of producing an invented complete total. Previously stored assessments retain their original evidence and rules; the upgrade does not reprice historical records. The catalog adds verified Sol Ultrafast Responses and Gemini Argon GenerateContent rates. Rules still require matching billing dimensions. The bundled catalog expires on November 8, 2026; existing verification dates were not extended. Use a reviewed pinned catalog or contract rules when its coverage does not match your account. See cost accounting.

​
Opt in to Gemini Argon

Select google("gemini-4-argon") explicitly. Existing saved model IDs are unchanged. Argon accepts thinking levels, but rejects disabled reasoning, none, minimal, token-based thinking budgets, explicit cached-content resources, and output limits above 65,536 before transport. Smaller output limits and caller timeouts are preserved. The adapter supports a single GenerateContent request. Long-decode orchestration is not included. A continuation-required response throws ModelUsageError with available usage and does not emit a successful finish. Treat earlier streamed text as provisional until finish. See Gemini Argon.

​
Retain signed tool history

When your application stores Google conversation history, retain the full assistant message, including providerExtras. Version 4.11 preserves native call IDs, signed function parts, and requested/served model identities across terminal stream chunks. Reconstructing only the visible text loses provider continuation evidence.