CHANGELOG
What's new in Nexora
Every improvement, fix, and feature we ship — updated continuously.
Guardrails 2.0
NEW- Policy-as-code guardrails now run inline with generation and block only on violations, adding under 8ms of latency.
- New per-tool scopes let you attach different policies to each tool a single agent can call.
- Added 12 prebuilt policy packs for PII, prompt injection, and toxicity — enable them from the dashboard.
- Fixed a race where streamed tokens could bypass a late-loading policy on cold agents.
Distributed tracing, everywhere
IMPROVED- One trace id now spans every sub-agent, tool call, and model hop in a single timeline.
- Token-level cost and latency are attributed to each span so you can find the expensive step fast.
- Traces export to Datadog and OpenTelemetry collectors with a one-line config.
- Improved the trace viewer to lazy-load deep runs — 10k-span traces open instantly now.
Faster inference routing
IMPROVED- Rewrote the router to keep warm capacity in every region, dropping p95 first-token latency to 82ms.
- Automatic failover now reroutes to the next-best model within a request when a provider degrades.
- Added weighted A/B routing so you can shift traffic between model versions without a redeploy.
- Fixed an edge case where retries could double-count usage on partially streamed responses.
Evaluations get rubric judges
NEW- Reference-free rubric judges score outputs against criteria you define in plain language.
- Evaluation runs can now gate deploys in CI — fail the build when quality drops below a threshold.
- Added drift detection that flags when live quality diverges from your last eval baseline.
- Judges run on the same warm capacity as inference, so eval suites finish in seconds.
The typed Agents SDK
MAJOR- Introduced the typed Agents SDK: define tools, memory, and structured outputs in code with full type safety.
- Streaming, retries, and multi-region deploys are handled by the platform — no orchestration glue.
- Shipped 40+ native connectors for the models, data stores, and tools you already use.
- New unified dashboard brings agents, inference, evals, and observability under one roof.