LoginSign Up
← Back to all sparks
E

ElevenLabs

INFRA · APIS
Velocity6.3

AI voice generation platform for creating realistic text-to-speech and voice cloning.

elevenlabs.io ↗

ElevenLabs is rebuilding voice agents as versioned software — branches, drafts, compiles, tests.

voice agentsagent versioningprocedurestelephonyknowledge baseagent testing
◆Current state
ElevenAgents now dominates the changelog, and it ships software-engineering primitives rather than voice features: branches, rebase, merge previews, drafts, and per-test tool mocks. Procedures compile into agent workflows, knowledge bases have crawl jobs and bulk operations, and telephony has widened past Twilio to Exotel and SIP. The speech and music models still advance — Music v2, Speech Engine — but they now read as components sitting under an agent runtime.
◆Where it's heading
The company is converging on an agent SDLC: author in a branch, hold changes as drafts, preview the merge, compile, then test against mocked tools. Every release adds another step of that loop rather than another voice. The analytics surface — sentiment scoring, termination reasons, backchannel flags, per-branch call volume — is being built to close the same loop with production evidence.
◆Prediction
With simulations, tool mocks and is_error failure paths now in place alongside branches and merge previews, the missing piece is a gate: expect evaluation runs that block a branch merge on regression rather than leaving testing as a manual step.

◆Recent moves

  1. 9d ago

    Per-test tool mocks and a knowledge-base system tool

    Testing gets sharper: tool_mock_overrides replace shared mocks for a single test, and an optional is_error flag lets a simulation drive the tool-failure path on purpose. This is the test-harness half of the agents-as-code arc, filled in behind the branch and draft machinery that arrived earlier.

  2. 16d ago

    Procedure APIs: draft, compile and branch agent behavior

    ⚡ SPARK

    Procedures become a branch-scoped authoring unit with their own draft-and-compile lifecycle, slotting into the version-control model the platform has been assembling since spring. Crawl jobs and bulk knowledge-base operations extend the same push to make agent content manageable at volume.

  3. 1mo ago

    Nested agent transfers gain push, pop and replace

    Transfers acquire a call-stack shape: push into a nested workflow, pop back to the parent, or replace outright, with explicit controls for whether the child returns. Sentiment scoring, transcript auto-translation and backchannel flags round out the post-call analytics surface.

  4. 1mo ago

    Merge and rebase previews for agent branches

    You can now inspect a merged or rebased agent configuration — including which fields get overridden and whether source and target already match — before committing it. Seven-day call volume per branch makes branches measurable rather than only editable.

  5. 1mo ago

    Branch rebase, plus product and termination filters

    Rebase moves a branch onto the latest main while keeping its own changes, and versions now record the source they were rebased from. The new conversation filters — product type, termination reason — are the reporting side of a call corpus that has grown large enough to need slicing.

  6. 1mo ago

    Introducing Music v2

    ⚡ SPARK

    Music v2 trades prompt-only generation for chunk-based composition plans, handing back control over structure and pacing. Against a changelog otherwise consumed by agent tooling, it signals the model line is still being advanced rather than parked.