Every response this agent sends has a declared source — a file, a function, a canonical message. Not logs. Not a black-box trace. The Console shows what it said, why it said it, and where the data came from. If something doesn't look right, you change it.
The price doesn't come from the model. It comes from the database, assembled by code, appended after the LLM has already run. If the model tries to write a number — even if the customer asks three times — a guard strips it before it goes out. The same applies to bank transfer data, product measurements, what's included, and every canonical link. 17 of 25 capabilities work this way: the output is identical, turn after turn, regardless of the model's state.
(01) — Human sovereignty
It's a design rule. When a human steps into a conversation, the agent doesn't respond on top of them. It reads the intervention, marks the thread as in-handoff, and goes quiet. No overlap. No race condition. The customer gets one voice at a time.
(02) — Deterministic where possible
Price, measurements, stock, payment terms — resolved from controlled sources, not from the model's best guess. The LLM enters only where it adds something a lookup can't: warmth, context, judgment. Everywhere else, code handles it.
(03) — Honest fallback
If a response doesn't pass the guard, the system doesn't hallucinate an answer. It degrades honestly, logs the state in the CRM, and leaves a full trace. The customer gets a coherent experience. Operations gets a signal. Nothing gets papered over.
(04) — Recovery loop
Outbound queued. Status confirmed. Retry on failure. A sweeper catches what the queue missed. Ops alert on terminal failure. The system is operable, not just intelligent — if something breaks, there's a path back.
(05) — Coexistence
No notification needed. When the operator replies from the app, the system detects it and silences the agent for 12 hours — the lead keeps its state, the CRM updates. They don’t compete, by design.
The 8 open responses are the ones that actually need context: recommending a product based on the customer's use case, handling a price objection without a script, recovering from a confused exchange. That's where the LLM enters — with the full conversation history, typed intent signals, and a hard ceiling on what actions it can take. The Console shows what triggers each one and which source it pulls from.
Price, measurements, product cards, links, escalation. Verbatim. No variation between turns.
🔒 CODEGreeting, recommendation, qualifying, objections, follow-up. Generated in real time.
⚡ LLMRepresentative triggers
Switch on X-ray and tap any bubble in the chat to see its spec — triggers, source in code, and the editable canonical message.
These phrases trigger this response. Code detects them — the LLM doesn't choose the path.
Exactly what the customer receives. It comes from code — the LLM never writes this block.
