CygnusAlpha

1. Repetitive, expensive supportMost tickets were routine status questions, handled one by one by human agents.
   →
   FixAI agents answer routine questions on every channel, including voice bots and automated follow-up calls.
2. Answers live in other systemsOrder status, balances and account details sit in CRMs and internal APIs, not in documents.
   →
   FixAgents read live data from each client's CRM and order systems.
3. Acting safely, not just talkingLetting AI change records is only acceptable if every action passes the rules a human would.
   →
   FixA short list of allowed actions, each checked against the client's business rules.
4. One platform, many clientsEach client needs its own data, prompts and limits on shared infrastructure that holds up under load.
   →
   FixOne multi-tenant core: per-client data, prompts and limits on an event-driven backbone.

Each request is handled by a small team of agents. Every agent gets only the tools its job needs, and only Action can write.

• Router

  Reads the request and picks the right agent.

  no tools, routes onlyNO TOOLS

  e.g. "refund" goes to the Action agent
• Knowledge

  Answers from the client's own content, with retrieval.

  search client contentREAD-ONLY

  e.g. A return-policy question
• Live data

  Reads the client's CRM and order systems, read-only.

  order lookupREAD-ONLYaccount lookupREAD-ONLY

  e.g. "Where is my order?"
• Action

  Takes only allowed actions, each checked by business rules.

  create ticketWRITE · RULE-CHECKEDupdate recordWRITE · RULE-CHECKED

  e.g. Raises a support ticket
• Voice

  Voice bots answer calls, and automated follow-up calls go out, on the same agents.

  speech in and outVOICEfollow-up callsVOICEsame tools as chatREAD-ONLY

  e.g. An automatic follow-up call
• Hand-off

  Passes hard cases to a person, with full context.

  human queueWITH CONTEXT

  e.g. A disputed refund

• AI that changes data. Every change is checked against the client's business rules.
• Production, not a demo. High load, many clients, strict uptime, all at once.
• Quality over time. Changes are evaluated locally before reaching a client.
• The trade-off. Two vector stores - Pinecone for the managed production path, ChromaDB for local iteration and evals - some duplication in exchange for not running vector infrastructure ourselves.

1. Give an assistant a short list of allowed actions, not open access.
2. Keep business rules where the model can't override them.
3. The cheapest regression is the one a client never sees.