Skip to content

October 2026 documentation updates

Found by the SDK dependency-compatibility run, which installs the newest framework releases each extra allows; shipped as Server and Python SDK 3.2.3.

  • SDK 3.2.3. Strands 1.62.0 moved its middleware package and added telemetry options to its tool terminal, and SDK 3.2.2 refused to build a client once strands was imported. The automatic Strands guard now covers the 1.62 terminal: authorization runs before the tool span opens, and a denied call opens none. On 3.2.2, pin strands-agents<1.62.
  • Server 3.2.3. Version metadata only.

23 October 2026: Team and Business prices revised

Section titled “23 October 2026: Team and Business prices revised”

The October 2026 price card is revised to bring the self-serve plans in line with the market entry points. Features, allowances, seats and retention on every plan are unchanged; only the two prices and one overage rate move.

  • Team is now $49 per month or $490 per year (was $99 / $990). Extra requests are $9 per 100,000 (was $12). 500,000 governed requests and 10 seats stay included; extra seats stay $10 per month, prorated.
  • Business is now $249 per month or $2,490 per year (was $399 / $3,990). Extra requests stay $6 per 100,000; 3,000,000 governed requests and 30 seats stay included.
  • Developer, Enterprise VPC and Enterprise Self-Hosted are unchanged.
  • The amount shown at checkout is the amount charged. Indian customers are invoiced in INR at the prevailing reference rate plus 18% GST.

23 October 2026: streaming errors keep their cause

Section titled “23 October 2026: streaming errors keep their cause”

Found by a measured run of the QA catalogue while the QA provider key’s OpenAI account had no credits; shipped as Server and Python SDK 3.2.2.

  • Streaming error relay. A provider error before the first chunk answers the real status with the format’s JSON envelope instead of 200 plus an empty error event; a mid-stream error carries the status, type and message in the OpenAI, Anthropic and Gemini formats, and the stream then ends.
  • SDK 3.2.2. A composite LangGraph node’s first execution attests the same composition the discovery reported, so it no longer quarantines the agent with blueprint_coverage_incomplete.
  • QA catalogue. [dev,all] extras for SDK source tests; user_data file cases check stored metadata; agentic cases send evaluable targets; runner tools for the guardrail-rule family and the automated catalogue.

22 October 2026: 3.2.1 corrections from the QA error review

Section titled “22 October 2026: 3.2.1 corrections from the QA error review”

Found by reviewing every error, failure and blocked case the internal QA testers had recorded against the deployed 3.2.1; shipped as Server and Python SDK 3.2.2.

  • Decision targets. POST /api/agentic-new/decide validates object (<type>:<id>, a model type, an id without : # * or whitespace) and permission (a relation of that type or a custom key containing :) before asking OpenFGA, and answers 400 invalid_object / 400 invalid_permission naming the value and the accepted forms, where the decision used to read authz_engine_error.
  • Blueprint scans. A capture with more than 192 static findings is blueprint_coverage_incomplete (400) instead of a 503 blueprint_scan_unavailable quarantine.
  • Guardrail evaluation answers with the additive policies_evaluated list, so a disabled policy is visible as absent.
  • Prefilter. Multipart uploads (/v1/files) and SDP bodies pass the JSON fast check, which now applies to JSON content types only.
  • Console. Provider-key disable persists for keys created without an explicit flag; policy and RAG source names are required with specific messages; RE2-incompatible custom regexes are named at Create card; numeric fields reject non-numeric and negative input inline; field-level API errors reach the form.
  • QA catalogue. Missing settings, a missing source checkout and workspace-refused models record Blocked with the exact item to fix; UTF-8 console output; MFA enrolment first; plan-gated pages Blocked; console cases reworded to the product’s actual controls.

18 October 2026: 3.2.1 corrections continued

Section titled “18 October 2026: 3.2.1 corrections continued”

Found while running the internal QA catalogue against the deployed 3.2.1; all in Server and Python SDK 3.2.1.

  • Gemini parameters_json_schema. A function declared with the google-genai SDK’s plain JSON Schema reached the provider without parameters, so the model called it with empty arguments; the schema now becomes the function’s parameters.
  • SDK: a bare AutoGen agent. An autogen_agentchat or AG2 agent run without a team is discovered as its own program with its model and its definition (the system message) as the code to scan, instead of a report with no artifact that read blueprint_coverage_incomplete and quarantined the agent.
  • Cloud Run tier. The deploy creates one guard shared secret in Secret Manager and mounts it into the server and models containers (the fine-tune page and scoped evaluations need it), and renders the durable webhook outbox on with its object-store settings.
  • QA catalogue. Fenced or prefaced JSON answers parse; API contracts of a flag-mounted feature read Blocked when the deployment does not mount it.

The SDK, server and documentation reference move to 3.2.1; see Server and Python SDK 3.2.1.

  • MCP tool schemas only for callers that can use them. A virtual key bound to MCP servers but, with Agentic Security on, to no enabled agent is no longer shown MCP tool schemas, so a plain application call through such a key completes instead of ending in 503 mcp_tool_authorization_unavailable at the model’s first tool call. Keys with an enabled agent are unchanged.
  • Unknown API paths. /api, /v1 and /scim paths that do not exist answer the JSON 404 envelope instead of the dashboard’s HTML.
  • Gemini forced function calls. A Gemini-format request that forces one named function converts to a tool_choice carrying that name, so providers that validate the choice no longer answer 400 Missing required parameter: 'tool_choice.name'.
  • Cost-optimization cascade and the model policy. The cascade asks the guardrails plugin before substituting a cheaper tier; a tier the caller’s model-inventory policy forbids is not used and the request keeps its own model, instead of being refused for a model it never asked for.
  • Operations. Redis 8 semantic-cache index deletion, and every plugin module builds and tests standalone.
  • SDK. Aligned at 3.2.1; APIs unchanged from 3.2.0.

11 October 2026: two-run registration, deferred activation and registration e-mails

Section titled “11 October 2026: two-run registration, deferred activation and registration e-mails”

Registering an agent now takes one developer run, one administrator sitting and one more developer run. The SDK, server and documentation reference move to 3.2.0; see Server and Python SDK 3.2.0.

  • Developer route. With SDK 3.2.0 or newer, the developer runs the agent with four settings - DEEPINTSHIELD_BASE_URL, DEEPINTSHIELD_VIRTUAL_KEY, DEEPINTSHIELD_AGENT_NAME and DEEPINTSHIELD_REGISTERED_BY (constructor argument registered_by) - and sees only Registration pending. Reference <registration id> · runtime key SHA256:<thumbprint>. Contact your administrator. You will be notified at <email>. There is no review link; the developer cannot approve. The closing sentence appears only when the gateway confirmed the address.
  • Registering e-mail. DEEPINTSHIELD_REGISTERED_BY is the explicit way to name the registering developer. The gateway maps it to a person in the workspace directory, created just in time when unknown, and notifies it. It is not verified and grants nothing: it is proposed as the accountable owner, which the administrator confirms or changes. One registration keeps its first address; a different address is refused with registration_requester_mismatch. When it is unset, an e-mail-shaped requester is only an advisory hint: the gateway uses it only for a person the organization placed in the workspace and ignores it otherwise, without refusing the run.
  • Runtime key on run 1. In Entra gateway mode the first run already submits the SDK’s runtime key, as a Provisional binding when the identity does not exist yet. The key goes to the workspace’s validated gateway-exchange Entra provider; Default for new registrations selects one when there are several. The key is authorized in the review; authorizing it before the identity exists is refused with runtime_binding_identity_missing.
  • One sitting. The administrator completes owner, identity (create or link it from the review), runtime key, code and tools in one visit of the registration review.
  • Deferred activation. Approve now, activates on the agent’s next run records the decision pinned to the source digest, scan, declaration revision and runtime-key binding; the next run of the unchanged script applies it before answering. A change marks the decision Stale decision · <reason>; Withdraw decision and AGENTIC_DEFERRED_DECISION_TTL_HOURS (default 168) close it otherwise. A recorded decision grants every tool action the reviewer selected; the Decision recorded card lists the actions that activation will grant, so the next run executes the whole agent. Resource permissions and delegations are added after activation. See Deferred activation.
  • Four-eyes and two settings. A deferred decision always needs a reviewer other than the accountable owner and other than an administrator who registered the agent with their own sign-in e-mail (approver_conflict_of_interest). The Registration review card on the Runtime policy tab of Policy & Access adds Require a separate reviewer for immediate approvals (default off) and Allow VK-trusted approvals (development only) (default on); the review option is now labelled Approve as VK-trusted (development only).
  • Every identity type. One table lists, per identity type, what the sitting contains, what the developer needs before run 2, how many runs it takes and whether server runs apply. Approve as VK-trusted (development only) is offered with or without a declared identity. Only an identity that uses gateway exchange can finish a server run. See the registration route by identity type.
  • Registration e-mails. The developer receives a pending receipt, the decision, approval and denial; the administrators receive a review-needed message with the review link; a recorded decision that becomes stale is reported to both. Messages are sent from the platform no-reply mailbox (SMTP_FROM_EMAIL) through the platform SMTP_* settings, with no separate configuration; they are written to a durable outbox in the same transaction as the decision and never on the authorization path. While SMTP is not configured they stay queued and are sent once it is. See Registration notifications.
  • SDK. The Entra direct chain tries environment variables, workload identity federation and managed identity once each and fails fast with credential_host_identity_missing; new Entra providers default to Gateway exchanges Entra token · recommended. A pending registration is re-sighted after agent_runtime_enrollment_pending, credential_exchange_failed and credential_host_identity_missing so the registration record stays current. See SDK error codes and configuration.
  • Deployment. The Cloud Run deployment script no longer reports the digest-resolved sidecar images of a candidate revision as configuration drift; a sidecar moved to another repository still stops the rollout.

This entry documents the source implementation; it does not announce a release. Check the version and enabled capabilities of your own gateway.

Revised 23 October 2026. Team is now $49 per month or $490 per year with extra requests at $9 per 100,000, and Business is $249 per month or $2,490 per year. Allowances, seats, retention and features are unchanged. See 23 October 2026: Team and Business prices revised.

Billing and monthly extras now documents the October 2026 price card. It replaces the rates introduced in September on every plan. Plans remain per organization and billed by calendar month (UTC), and annual billing remains ten times the monthly price. There is one worldwide price list in US dollars; Indian customers are invoiced in INR at the prevailing reference rate plus 18% GST, and the amount shown at checkout is the amount charged.

  • Developer stays free with 50,000 governed requests and 2 seats included each month. Both are hard caps: once the allowance is used, further governed requests are rejected with HTTP 429 (plan_request_allowance_exceeded) until the month resets, and a third invitation is refused with HTTP 402 until the organization upgrades.
  • Team is $99 per month or $990 per year with 500,000 governed requests and 10 seats included. Extra requests are $12 per 100,000 and extra seats $10 per month, prorated. There is no hard cap; Business is recommended past 3,000,000 requests a month.
  • Business is $399 per month or $3,990 per year with 3,000,000 governed requests and 30 seats included. Extra requests are $6 per 100,000 and extra seats $10 per month, prorated. There is no hard cap; talk to us about Enterprise past 10,000,000 requests a month.
  • Enterprise VPC lists from $12,000 per year (contracted) with 10,000,000 governed requests a month, 50 console seats, 500 protected identities, and one production data plane. Enterprise Self-Hosted lists from $24,000 per year (contracted) with 25,000,000 governed requests a month, 50 console seats, 1,000 protected identities, up to three production clusters, and an air-gapped delivery option. Enterprise overage is graduated on the excess and metered on aggregate counters: $5 per 100,000 requests for the first 20,000,000 over, $3 per 100,000 for the next 75,000,000, and $2 per 100,000 beyond; each additional protected identity is $0.60 per month for the next 1,000, $0.40 for the next 4,000, $0.25 for the next 10,000, and $0.15 beyond.
  • The unpaid-extras credit limit stays at twice the monthly base including tax, which is now Team $198 and Business $798 by default.

Log retention, virtual-key, workspace, MCP-server, guardrail-policy, and routing-rule limits, every feature entitlement, the 0% token markup, UPI AutoPay, bank transfer, and the collection timeline are unchanged. This entry documents the price card in the embedded catalog; the rates your deployment charges come from its configured catalog, and the September notes retain the rates they describe.

10 October 2026: redesigned console and current screenshots

Section titled “10 October 2026: redesigned console and current screenshots”

Every product screenshot on this site and in the quick start now shows the redesigned console: navigation grouped under ACTIVITY, GOVERNANCE, RUNTIME, and RESOURCES, one merged tenant and workspace switcher, the Search or jump to… field, the plan row at the bottom of the menu, and Continue setup, the theme toggle, and the documentation link in the top bar. The captures use synthetic demonstration data only.

Page text follows the current labels and paths, for example Model Hub → AI Providers for provider keys, Access & Credentials → Virtual API Keys, Analytics → Overview for the dashboard tabs, and Agentic → Agents → Needs attention with its Review queue. Steps that named retired controls were rewritten against the current interface.

10 October 2026: server runs on every workspace and repeat runs

Section titled “10 October 2026: server runs on every workspace and repeat runs”

The managed-execution guide, SDK reference, Agentic guide, deployment overview, and quick start now describe the current server-run behavior:

  • Every workspace: an operator can enable server runs for every workspace at once. Limiting them to selected workspaces is optional. The managed-execution guide explains when a workspace is offered a server run.
  • Any client version: the worker runs its own SDK, Python and package releases and checks only that each package the program imports is installed. Client versions are recorded as review evidence and never need to match.
  • Repeat runs: with a gateway and SDK that support submission IDs, running the unchanged script again starts a new server run under the existing approved registration. An active or uncertain run must be resolved first.
  • Worker runtime: the executor is either an operator-managed Docker host or, on platforms without Docker such as Cloud Run and Kubernetes, a separately deployed worker service that the gateway calls over HTTPS. Gateway hosting alone supplies neither.
  • Review: a waiting server run shows Server runtime unavailable when the worker runtime cannot be reached, and Runtime access states the next step for a workload that brings its own runtime key in a workspace without server runs.

This entry documents the source implementation; it does not announce a release or deploy a gateway or worker.

8 October 2026: 3.1.0 and the current agent workspace

Section titled “8 October 2026: 3.1.0 and the current agent workspace”

The documentation reference now targets Server 3.1.0 and Python SDK 3.1.0. Current install examples, integration guides, and SDK error tables match that release. Historical notes retain the versions they describe.

Agents consolidates Needs attention, Agent inventory, Denied agents, Discovery, and Shared assets. MCP Registry owns live connections and settings. Registration review follows the six-section checklist and links identity, exact source/tool access, runtime readiness, and the final decision. The new managed-execution guide explains saved LangGraph submissions and the optional server worker.

The header, favicon, and social preview use the official DeepIntShield brand. Updated interface screenshots show the current implementation using synthetic demonstration data. Read 3.1.0 release notes for upgrade steps and the limits of managed execution.

4 October 2026: Microsoft Entra and local gateway walkthrough

Section titled “4 October 2026: Microsoft Entra and local gateway walkthrough”

The new complete local setup takes a fresh Microsoft tenant configuration through a local gateway and math-agent in test2.py: gateway API audience/role, Agent ID blueprint, separate management app and consent, encrypted pasted credentials, opt-in automatic Gateway.Invoke assignment, runtime-key authorization, Entra proof and code/access review. Fresh startup includes persistent local OpenFGA/PostgreSQL dependencies and requires successful authorization initialization. It includes a sanitized Python example, expected pending states, troubleshooting, setup frequency and hosted/cloud follow-on guidance. Seven Microsoft Learn portal screenshots now illustrate registration, API identifier, application role, blueprint creation/credentials, API permissions and admin consent. Each is preserved unchanged with source attribution, Microsoft’s MIT license and a caption explaining which example selections differ from this setup.

The gateway quickstart, SDK reference, SSO/workload guides and homepage now link to that flow. They distinguish direct workload credentials from the gateway’s blueprint credential and the SDK’s approved runtime key. Manual credential environment references and encryption-key setup are not required for fresh pasted Entra credentials. This documents the source implementation; it does not publish SDK packages, deploy a gateway or claim live tenant/cloud validation.

Source capture now distinguishes SDK instrumentation from application code when an editable SDK or virtual environment is nested under the script’s directory. Application decorators and helper changes remain included in code review. Shared extraction covers the callable fields used by supported agent frameworks instead of assuming every tool stores its implementation as func.

Native Google GenAI clients returned by shield.genai() preserve their model, chat, async, and streaming APIs while governing automatic Python callable-tool execution. Governance failures stop the function call and reach the caller; they are not converted into an ordinary tool error for the model to retry. See Agentic SDK integration.

4 October 2026: Complete Agentic operator workflow

Section titled “4 October 2026: Complete Agentic operator workflow”

The Agentic guides and website quick-start now follow the current Work Queue, Assets, Identities, Policy & Access, and Activity navigation. They cover initial discovery and registration, provider and provisioning setup, source capture and code review, exact tool/action and resource access, runtime and task policies, temporary delegation, action approvals, enrolment, schema/code drift, and evidence review.

The provisioning walkthrough includes server authorization references, Entra sponsors and workspace link authorization, connection checks, create/link jobs, automatic expected-identity declaration, and the separate workload-proof step. A completed provisioning job does not approve code or tool access.

MCP setup now points to Assets → MCP connections, including its Connections and Settings tabs. Activity documentation includes the Audit tab’s recent 200-change window alongside Runs, Decisions, and Trends. Guided access availability follows the current Business entitlement; Team retains its included context-policy and historical activity views.

See Agentic Security, workload identity, the SDK walkthrough, and Agentic observability. This documents the current implementation and regenerates the website content; it does not publish a release or deploy a gateway.

The Identities → Users and Service accounts tables show the newest created entries first. After a successful manual creation, search and source filters clear and the table returns to page one so the new entry is visible. Page size adapts to the viewport, with at least two rows per page when available; short screens can scroll to the remaining rows and pagination controls.

The directory API applies the optional order=recent sort before pagination, so recent entries are not lost outside an alphabetical result window. Calls without that option keep their existing order. See the identity setup guidance and directory API boundary.

4 October 2026: Workload identity provisioning guidance

Section titled “4 October 2026: Workload identity provisioning guidance”

The guides now distinguish Agentic → Identities → Provisioning from runtime Identity providers. They cover Entra Agent ID, Keycloak, Auth0, and Okta; customer-approved management references; stable identities across agent runs; and the supported lifecycle actions for managed and read-only linked identities. Identity creation leaves workload proof, code review, and access approval intact.

Authorize provisioning reuses the chosen verifier’s issuer, audience, and Entra tenant/blueprint as read-only settings. New shared Entra verifiers use an explicit empty identity list, allowing validation before the first child exists while admitting no workload. Provisioning fills the list; existing admission policies and single-child configurations are preserved.

The provisioning view separates Created & linked records from Existing identity declarations. A zero provisioning count does not discard a prior declaration or its proof history. An unused connection can explicitly apply changed verifier settings with Use current verification settings, then save and check the connection. Used connections keep their original trust settings and require a new connection for a different configuration.

Provider edits now persist the current admission list. Shared Entra verification and credential setup honor each agent’s declared child when no default child is set; an empty shared list continues to reject all workloads. These fixes do not widen saved admission policies or bypass proof and approval.

Vault delivery guidance requires ready: true and separate workload access to the secret store. An SDK launcher example consumes a mounted ready record without printing credentials. The SDK’s new OIDC_SCOPES override uses per-client scopes, falls back to provider defaults when unset, and requests no scopes when empty. Use an SDK build containing that override. Recovery guidance covers durable job retries, local quarantine before disable/retire, safe release of unused declarations, and re-enrolment that retains prior proof and denial history.

This entry documents the implementation and examples; it does not announce a package release, deployment, or live validation against customer tenants. See workload provisioning, SDK credential setup, and enrolment recovery. The gateway quickstart and SDK overview now point new users to the same governed-agent setup without repeating the advanced configuration.

4 October 2026: Agentic onboarding and runtime drift guidance

Section titled “4 October 2026: Agentic onboarding and runtime drift guidance”

The current guides now show the single Confirm approval and access flow for registration, the completed current blueprint, and selected action/resource grants. The simple-math-agent LangGraph walkthrough covers both discovered invoke actions, readiness, and repeat runs. Expected identity declarations, verified workload proof, and explicit Approve as VK-trusted remain distinct.

Block runtime drift (blueprint_runtime_enforce) defaults off. Mandatory scanning remains active; strict mode quarantines every newly observed different digest for review, including clean changes and rollback to an older approved digest. Turning it off does not release an existing quarantine. Coverage is limited to reported code and gateway-observed tool definitions before subsequent protected actions. It does not monitor arbitrary host changes or cancel actions already running.

The protected-key model picker uses the selected key’s scoped provider/model configuration and filtered catalog, clears incompatible models, and rejects stale or failed loading results. The server still validates saved routes.

These are documentation updates for the corresponding server/UI changes, not a release or deployment announcement. No SDK changes are required for these flows or strict drift protection; publishing documentation does not deploy or validate them in production. See Agentic, the LangGraph walkthrough, workload identity, and enrolment.

2 October 2026: guardrail configuration guidance

Section titled “2 October 2026: guardrail configuration guidance”

The public documentation now covers the latest policy-builder and pattern configuration changes. This is a documentation update, not a new SDK or server version announcement. The described editor and validation behavior requires a gateway deployment containing these changes; publishing this site does not deploy them.

  • Policy-card editing: use Settings for enablement, failure action, severity, summary, redaction replacement, and preset selections. Use Patterns / JSON to inspect generated checks or enable a card override.
  • Pattern overrides: an override replaces the card’s generated detectors and lists. The card’s Settings retain control of action, severity, summary, and enabled state. Per-check scopes can narrow the card’s supported stages.
  • Saving and publishing: Save details updates the policy draft; Save Policy saves and publishes an existing policy. Create Policy creates a draft that must be published explicitly. Reset to generated settings removes an override. Review the effective configuration and test both matching and allowed examples before publishing.
  • Regex validation: policy management validates expressions using the runtime’s Go/RE2 parser. Lookaround and backreferences are unsupported. Invalid configurations return errors when policies are created, updated, published, or rolled back.
  • Preset catalog: the new catalog guide explains coverage, examples, and limitations. Authoritative version 2 policy checks retain their saved configuration; legacy builder policies regenerate against the installed catalog and should be reviewed and republished to pin their checks.

See Guardrails and the guardrail pattern catalog.

The homepage and changelog also now identify Python SDK 3.0.0 separately from the Server 2.8.3 compatibility reference, with the RAG migration and redaction-decision notes collected in one release entry.