2.8.3

DeepIntShield Server and Python SDK 2.8.3 — 10 September 2026. Check the version reported by your gateway separately from the installed SDK; these notes describe the release and do not establish a hosted deployment’s version.

Agentic execution and workspace controls

• Developer and Team workflows execute without entering Agentic Security. They no longer fail on an unavailable Agentic feature, pending registration, or blueprint review. Work Queue and Guided access show Upgrade to Business for these plans.
• On Business, Enterprise, and Enterprise VPC, workspace administrators can turn Agentic Security off or on from Agentic → Work Queue. Existing workspaces default to on. Disabling skips Agentic discovery, blueprints, identity proof, decisions, approvals, and run-ledger calls; re-enabling restores governance. Stored policies, registrations, and historical activity remain available.
• The SDK reads the authenticated workspace runtime setting for each new invocation. Invalid responses, authentication failures, and unavailable settings do not disable governance. A gateway without the runtime-config route retains the existing governed behavior until the server is upgraded.
• Gateway authentication, workspace boundaries, and Virtual Key MCP restrictions remain enforced when Agentic Security is off. A disabled run does not create synthetic governed activity or analytics records.
• Protected console pages wait for workspace and entitlement resolution. A retained identity feature override on a downgraded Developer or Team plan cannot unlock the Work Queue.

See Agentic plans and controls and SDK governance.

SDK compatibility and inference

• Removed the CrewAI dependency extra and aggregate dependency, native binders, automatic framework integration, and advertised server framework support. Remove crewai from DeepIntShield extras and migrate calls to a supported framework or inference client before upgrading. Generic callable tools retain their explicit governance APIs.
• LiteLLM completion() and acompletion(), including streaming calls, use the gateway’s inference path without synthetic Agentic blueprint or llm.completion checks. This also holds inside an Agentic run. Actual tool and workflow boundaries, and explicit Agentic decisions, retain their governance behavior.
• Updated the Strands terminal tool hook for the current call signature while preserving native tool context, native tool guards, and fail-closed Agentic decisions before execution.
• MCP tool-call and discovery rejections now retain their operation-specific error codes and safe numeric JSON-RPC details through the native session. Malformed responses remain protocol errors; canonical authorization and approval codes remain intact, including inside numeric JSON-RPC envelopes. MCP examples follow discovery pagination and check that the named tool is available before calling it.
• Gateway MCP catalogs refresh when authorized tools arrive, are removed, or change schema or description. Unchanged catalogs remain cached, and Virtual Key grants continue to restrict the exposed tools.

See LiteLLM and supported providers and frameworks.

Playground, analytics, and cache settings

• Playground input/output checks no longer require MCP or RAG entitlements solely because a shared policy also declares those scopes. Actual MCP/RAG evaluations and explicitly paid checks retain their existing entitlements.
• Agentic Analytics shows request failures with a retry action instead of displaying zero-valued results. Workspace and date changes discard stale results. Team retains its included historical analytics; Developer sees the appropriate upgrade prompt.
• Agentic Cache pages include persistent navigation to Settings, and show loading or retry states when configuration cannot be read. Switching workspaces resets the settings form to the current scope.
• Open workspace cache configuration at Agentic → Activity → Secure caches → Settings (/workspace/agentic-cache/settings). Individual cache controls remain under Agentic caches. Agentic Cache still requires Business or above.

See Agentic Cache.

Upgrade

Install the SDK with the extras your application uses:

python -m pip install --upgrade "deepintshield==2.8.3"
# Example for a native LangGraph workflow:
python -m pip install --upgrade "deepintshield[langgraph]==2.8.3"
python -m pip check

Upgrade the gateway to use the new plan-aware runtime setting and workspace switch. Review the dependency combinations before combining extras, and verify your selected framework’s tool execution and policy-denial paths. Native PydanticAI agent runs remain governed when Agentic Security is enabled.