Most voice platforms ship one giant LLM and hope it stays on script. We compose voice agents from small, testable sub-agents — each one a linear sequence of primitives.
Below: the thesis, the primitives, and what it looks like for recruiter screening.
ElevenLabs, Vapi, Bland start from one capable agent you steer with a prompt, then add workflows and tools on top. We start from the other end — a library of small, specialized sub-agents that compose into the call.
An orchestrator picks the right ones for each role. Cross-cutting listeners run across the whole call to score things like communication skill.
Voice agents in Feliciti are built like Notion documents — slash commands compose into blocks, blocks compose into sub-agents, sub-agents compose into the full call flow.
Nine primitives are all you need — one runs before the call connects, the rest run during.
Match the customer, fetch their record, load context — all before the first hello.
One bounded utterance. Variables interpolate, tone is set per sub-agent.
A scripted question that records a typed value back into the call state.
Asks only if a condition holds. The branch is explicit, not a prompt heuristic.
Commit a value to the DMS, ATS, or TMS. Triggers fire on side-effects.
Wrap up cleanly with a final line and end the session.
Warm-transfer to the right queue or advisor when escalation is needed.
When a step needs a small LLM call — kept narrow, kept testable.
The orchestrator's decision point — composition, not a chain of prompts.
A sub-agent is a named, reusable unit — a linear sequence of these primitives. Tested in isolation against its own scenarios. Composed with other sub-agents to form the full call flow.
ElevenLabs, Vapi, and Bland are excellent general-purpose voice platforms — they have workflows, branching, and tool-calling. The difference is what you assemble yourself versus what comes built for the work.
Earlier this year we needed a senior backend engineer.
We sourced tens of candidates through a technical recruiting marketplace.
Strong on paper. But we still needed to screen each one.
So we composed a screening agent the way we describe above.
An orchestrator picked the right sub-agents for the role:
verify_identity — to confirm we were talking to the right person.ask_seniority_pedigree — to probe their deepest project.ask_realtime_systems — because we needed voice agent experience specifically.ask_mutual_fit & ask_why_us — to qualify intent.kind_wrap — to close the call.The orchestrator skipped sub-agents that weren’t relevant per candidate.
As a founder, 20 intro calls × 30 mins is a full day lost per week.
That day could go into R&D, GTM, or anywhere that compounds.
Two candidates are hired and shipping in production today.
The rest got kind, structured wrap-ups in under 8 minutes.
If we’d shipped this as one giant prompt to a monolithic LLM platform, we’d still be tuning it.
Tell us the call you want to automate — recruiting or otherwise. We'll send back a working prototype within 48 hours, wired to your actual use case. Same composition pattern you just read about.