Changelog

Agent SDK 2.0.0 — every framework adapter published

rends-agent-sdk 2.0.0 (PyPI) and @rends/agent-sdk 2.0.0 (npm) are published. Python has adapters for LangChain, LangGraph, CrewAI, AutoGen, OpenAI Agents, Google ADK, Haystack, LlamaIndex and PydanticAI; TypeScript for LangChain, OpenAI Agents and Google ADK; both have an OpenTelemetry adapter. Each adapter was tested against its framework's current release before publishing. New in 2.0.0: an action held for human approval is reported as pending with its decision id, and resume() runs it only after approval and a policy re-check; tool calls from a paused or terminated agent are refused in enforce and monitor mode (checked against production on October 30, 2026); a response without a decision no longer acts. 2.0.0 is a major release — the SDK changelogs list what changes for 1.x users. Install: pip install rends-agent-sdk==2.0.0 or npm install @rends/agent-sdk@2.0.0.

Policy anchoring — verifiable decision records

Every evaluation now records the exact policy revision in force. Policy revisions are immutable and content-addressed by SHA-512 hash; each decision record pins both the policy ID and the revision hash. Anyone holding the audit chain can independently verify which rules produced a given outcome, without trusting Rends to retell the story.

Correction, September 29, 2026: This entry overstated what shipped. Revision stamps were recorded on some evaluations, not every evaluation, and audit-trail rows did not carry them, so the audit chain alone could not show which rule versions produced an outcome. Since September 28, 2026, every decision record pins the policy in force: a SHA-512 hash of the whole rule set in scope and, for each rule that applied, its exact revision and content hash. Policy anchoring, as it stands

Four-engine reframe — Discover · Translate · Govern · Comply

Translate joins Discover, Govern, and Comply as a first-class engine. The platform now positions as: where AI regulation becomes runtime control. Homepage, footer, navigation, and docs updated to reflect the four-engine architecture.

Translate engine — regulations to enforceable policy

Generates structured policy drafts from regulations using 7 control templates (human_oversight_gate, audit_log_required, data_classification_filter, model_allowlist, rate_limit_cap, output_content_policy, explainability_mandate). Approve a draft and the deploy bridge writes the controls into govern.access_policies in a single transaction; a Redis pub/sub signal reloads every govern instance's enforcement mesh. ~3¢ per draft on Haiku.

Custom regulations

Author org-private regulations alongside the 51+ public ones — internal AI policies, contractual requirements, board mandates. Enqueues the same translate pipeline as Radar-sourced regulations. CRUD under /v1/custom-regulations.

Scanner Lite — public AI exposure scan

Free, public-facing scanner at /scanner. Visitors describe their AI usage; an LLM pass returns a structured exposure summary in ~8 seconds. No account required, per-IP rate-limited and per-day spend-capped.

Python SDK — 4 new framework adapters

rends-agent-sdk now ships adapters for Haystack, LlamaIndex, PydanticAI, and LangGraph in addition to the existing LangChain, CrewAI, AutoGen, OpenAI Agents, and Google ADK. Ten framework adapters total — 4 on Node (@rends/agent-sdk: LangChain, OpenAI Agents, Google ADK, OTel), 10 on Python (Node set plus CrewAI, AutoGen, Haystack, LlamaIndex, PydanticAI, LangGraph).

Correction, October 5, 2026: These adapters are in the SDK source and are not yet in a published release. The published packages contain LangChain, CrewAI and AutoGen on Python (rends-agent-sdk 1.0.0) and LangChain on Node (@rends/agent-sdk 1.0.1).

Correction (October 29, 2026): the Haystack, LlamaIndex, PydanticAI and LangGraph adapters are in SDK 2.0.0, which is not yet published. Published today: Python 1.0.0 (LangChain, CrewAI, AutoGen), TypeScript 1.0.1 (LangChain).

Correction (October 30, 2026): these adapters are now published, in SDK 2.0.0: Python rends-agent-sdk 2.0.0 (all ten above) and TypeScript @rends/agent-sdk 2.0.0 (LangChain, OpenAI Agents, Google ADK, OpenTelemetry).

Multi-agent orchestration

Interaction chains, chain policies (max_hops, no_cycles, forbidden_pairs, target_class_blocklist), cycle detection, and a topology view for the agent-to-agent call graph. SDKs now forward chain_id and metadata.task_id automatically.

Guardian agent

Evaluates on its schedule to build behavioral baselines and detect rate, duration, token, semantic, and scope drift. Anomalies on adjacent agents correlate into a single incident, and every alert ships with an AI-narrated explanation.

Chain conflict detection

Hops sharing a task_id are paired and compared via embedding cosine distance. Drift becomes an advisory; large divergence opens a conflict that can block the downstream hop or escalate to review.

Approvals for chain challenges

A new decision='challenge' outcome pauses a chain hop, opens an approval request through the existing HITL queue, and resumes or blocks the hop based on the reviewer decision. TTL is configurable per chain policy.

Light enterprise design overhaul

Refreshed marketing and product surfaces with a lighter enterprise palette, clearer hierarchy, and more consistent spacing across the platform.

Multi-model adapter framework

Pluggable model adapters with a shared interface — swap providers without rewriting agent logic, with registry-backed configuration.

Three-engine architecture live

Deploy, Govern, and Comply now run as independent services behind the gateway, with clear boundaries and event-driven coordination.

Platform launch

Initial release of Rends — unified gateway, tenant-aware APIs, and the foundation for governed agent operations at scale.

Discovery Engine — 4 channels

Discovery Engine ships with four channels to find and classify AI agents across your estate — from repos to cloud spend signals.