Building Voice Agents That Last Some Lessons Learned From Forward Deployed Engineering

Treating an agent as deterministic software plus a drifting orchestrator changes how you version, test and release it. The post covers governance for agents that take actions in CRM and order systems, with a Revolut deployment as the example.

• The post splits a voice agent into conventional software (versioning, A/B tests, telephony config), which barely drifts after release, and a core orchestrator over language and audio, which can pass hundreds of tests and still fail in production.
• ElevenLabs recommends writing success criteria and agent tests before building, TDD-style, so go-live has a concrete bar: consistent passes across runs plus stable task-completion and escalation rates.
• According to the author, most pre-launch failures are prompt failures in three patterns: ambiguous instructions, tool misuse, and escalation drift, usually fixed by tightening an instruction, adding an example, or adjusting the escalation threshold.
• The post warns that batches under 100 calls per rollout branch are too noisy to judge: a 60% pass rate on 25 calls is much weaker evidence than the same rate on 100.
• Its reference targets: strong deployments with good data access aim for task completion above 80% and escalation below 20%, while teams with limited data access may find escalation near 50% a realistic ceiling.

• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

• articleIntroducing References Sound Control For Music V2
• articleEngineering Rag
• articleIntroducing Ivr Phone Tree Navigation
• articleFinetunes Music Api

• videoForward Deployed Engineering at DecagonAI Engineer
• articleIntroducing VersioningElevenLabs editorial sitemap
• articleBuilding An Agent For Our Own DocsElevenLabs editorial sitemap

Checking sign-in…

Loading comments…