Release Notes

AgentWatch’s current release line is 2.x. This page summarizes changes at a product level; contact your Iterate.ai representative for the detailed per-build changelog for your deployment.

AgentWatch 2.1 — highlights

Observe workspace. A unified observability workspace built into the console, spanning nine surfaces: AI usage and cost, service traces, fleet health, optimization comparison, business funnel, security threats, insider risk, delivery verification, and log analytics. The console is now fully themeable light and dark.

AgentFleet. The agent gains a sidecar and per-node mode for Kubernetes pods and Linux VMs, exposing explicit opt-in proxy endpoints rather than intercepting traffic. It reports Kubernetes node, pod, and namespace context, fails open so a workload never blocks on the gateway, and appears in the console’s fleet health grid. Ships with a multi-architecture image, DaemonSet manifest, sidecar-injection example, Helm chart, and a one-line VM installer.

AgentKeel optimization. Deployed agents can enable an in-process optimization stage before egress — exact and semantic response caching (a hit skips the provider call entirely) plus prompt and token compression for large repeated payloads. Off by default, opt-in per agent, and fail-open. The console reports model calls avoided, tokens saved, cache hit rate, and estimated cost saved.

Insider Risk for AI agents. Risk indicators across exfiltration, manipulation, boundary, and anomaly categories, computed live from DLP events, tool-approval decisions, and usage — with custom indicator authoring, risk-scored actors, data-access visibility, incident case management with an evidence timeline, and a watchlist. Adaptive enforcement maps risk score to warn, step-up, or block thresholds, applied inline in the gateway so a risky agent is stopped before reaching the provider. Enforcement is off by default.

Demo mode across every surface. All new surfaces return realistic generated data when demo mode is on, so an evaluator sees a fully populated console immediately.

Upgrading to 2.1 requires no schema migration — it is a plain image swap keeping the data volume.

AgentWatch 2.0 — highlights

Identity, substantially expanded. Native SAML 2.0 as a service provider, LDAP/Active Directory with LDAPS and CA certificate upload, SCIM v2 per-organization provisioning with managed tokens, and a unified hub-and-spoke setup wizard covering OIDC, SAML, and LDAP. Group-to-team mapping is unified across all three protocols, with off, automatic, and mapped modes.

Cross-protocol translation. OpenAI, Google Vertex AI, Azure OpenAI, and Cohere V2 inbound SDK formats route to any of the 14 upstream providers, with request and response bodies translated in both directions — so providers can be swapped without touching application code.

Endpoint agent for shadow AI. Signed installers for macOS, Windows, and Linux that capture and govern unmanaged public AI usage on managed devices without requiring a corporate proxy, with mutual-TLS enrollment, fleet management, account identity policy with dry-run testing and exemptions, and modern transport support including HTTP/3.

Transparent proxy mode. Domain mappings plus trusted-network validation let existing applications route through AgentWatch with no code change at all, while retaining full DLP, audit, budget enforcement, and per-employee attribution.

AI Insights. A natural-language query interface over your own audit and usage data, with persistent multi-turn conversations, trend and anomaly detection, and the option to run it against a local model so analysis stays in-boundary.

Per-employee governance. Every employee is a first-class identity with a scoped key, personal budgets, pause and resume, forced key rotation, CSV bulk import, and automatic task categorization of their requests.

Platform and reliability. Boot-time schema application and backfill so upgrades need no manual migration; self-healing that reaps stale jobs and re-probes tripped circuit breakers; zero-downtime rolling upgrades behind a load balancer; storage architecture reworked so large request bodies do not bloat the database; and encrypted backup and restore.

Operational notes

Upgrades are an image swap that preserves the data volume — do not remove volumes when bringing the stack down. Verify after upgrade by checking the health endpoint for the new version and confirming container health. Both insider-risk enforcement and agent optimization are off by default, so an upgrade does not change traffic behavior until a policy is explicitly configured.

Sources and references

• Product page: iterate.ai — AgentWatch
• Capabilities in depth: the Capabilities section in this documentation.