AI Agent = Model + Harness. We rent the model. We own the harness.
A strong model and a good runtime, by themselves, produce a poor selling agent — we learned that the expensive way. What turns them into a reliable one is everything around the model: exact commercial knowledge, an engineered and versioned system prompt, gated commercial actions, evaluation, and measurement. That layer is the Symphony42 Harness. It is where conversion is won or lost, and it is ours.
We deliberately rent the two layers below it. Frontier laboratories and voice platforms have spent billions on intelligence and modality delivery; we test each new release as a challenger against the incumbent champion, on conversion and economics, inside a harness that already works. That is META orchestration — and it is why every frontier-model upgrade makes Symphony42 sell better.
What we rent
Frontier speech models — OpenAI realtime speech-to-speech models · Voice agent runtime — Retell AI · Telephony and SMS — Twilio · Avatar and video — HeyGen LiveAvatar. Named challengers are evaluated continuously; a challenger is a candidate to test, not a promise to switch.
What we own — the Harness
Context Injection Service (CIS) — patent pending
A compact behavioral prompt governs how the agent sells. When the agent needs a fact — product information, pricing, objection guidance, competitor positioning — it calls CIS, which returns a small, speech-ready playbook by deterministic exact lookup: no vector search, no second model, the same answer every time. 100% accurate prices. 100% accurate product facts. Your knowledge is versioned and updated without touching the agent’s behavior.
Example: a consumer asks what the plan costs for a family of four. The agent calls the pricing tool; CIS returns your current pricing document; the agent explains it in plain speech while the authoritative price binds to the quote it will confirm. If the consumer says a competitor is cheaper, the competitor-intelligence tool returns your approved positioning. The agent stays fluent; the facts stay exact.
System Prompt Engineering
Every deployed agent begins with an approved Agent Spec — what it must accomplish, the claims it may and may not make, the exact language it must use, the trades it may offer — compiled into one complete, versioned prompt per deployment. Our prompt-engineering doctrine treats the system prompt as an engineered, tested, versioned asset, and binds everything that should never depend on generated language — transfer destinations, transaction values, calculations — in software instead.
Commercial Agent Tools
Real-time buyer match. Live transfer that preserves the caller’s number. Appointment setting. SMS quote and checkout links. Quote and CRM hand-offs. A clean, courteous end for non-fits. An action gate checks every precondition before a state-changing tool runs, and the quote-match invariant guarantees that what the consumer agreed to is what gets sent.
Evals — S42-TestCaller
AI-to-AI test calls against any Symphony42 agent, from ten scenario templates — the happy path, price objections, skeptics, ineligible callers, ambiguous intent, compliance probes, and more — before and after every release, with the evidence from both sides of the call.
Overture — the interaction layer
A click deserves better than a static landing page. Overture presents one clear primary action, one companion, and one alternative path — call, chat, text, browser voice, or video — and carries the consumer’s state across channels, so a person who starts in voice and finishes by text is one journey, not two.
Score — governed measurement
A semantic dictionary, a gate library, a metric registry, and immutable daily Recordings turn raw platform data into reproducible results. Every claim about what an agent did is a claim Score can check — and every outcome it records becomes proprietary evidence for the next improvement.
Runs on
Microsoft Azure (multi-region) · PostgreSQL · Redis · Replit · Cloudflare · GitHub. Control plane: Symphony42 OS, built on Claude.
Compliance and privacy by design
- Inbound-only, consumer-initiated: every Symphony42 call is dialed by the consumer or bridged by a partner that owns dialing compliance.
- Carrier-verified A2P 10DLC SMS for quotes and checkout links; consumers opt in on the call.
- No payment details by voice — ever. Assisted completion moves payment entry to a screen.
- No personal data in URLs; short-lived server-to-server hand-off tokens instead.
- Your data stays yours: never shared with third parties without your consent; CRM write-back configured per customer.
- Hard-coded compliance language and eligibility screens in the Agent Spec, enforced in software, not left to the model’s discretion.