Bland AI vs vLLM

Bland posts a dated changelog every two to three weeks, and the recent run is concentrated on making voice agents dependable rather than more impressive. Evals arrived in May under the named Sentinel release, agent testing and simulations plus CRM memory sync in July, and the newest entry adds adaptive resumption and node-scoped interruptibility — control over when an agent can be cut off mid-utterance and how it picks up afterward. The channel surface widened alongside it, with iMessage joining voice and SMS for enterprise accounts.

The arc runs from capability to control. Almost everything shipped since May either measures agent behaviour — evals, testing, simulations — or constrains it, through speech timing controls, per-node interruptibility, and scheduling status routing. That is the shape a platform takes when its customers move from pilots to production call volume and start caring about the worst call rather than the best demo. The plumbing releases point the same way: SIP outbound DIDs and full REST support for custom API tools are what an enterprise asks for before it routes real traffic through you.

Expect the eval and simulation tooling to keep deepening, most plausibly toward regression suites built from production call transcripts. The entries say too little about the CRM memory sync to tell whether it becomes a general memory layer or stays a per-integration feature.