The assistant factory that cites its sources
How Autoagentix Chat reads a business's system four ways, cites every fact to file and line, and hands over a grounded assistant the business owns.
The usual way a business gets an AI assistant is a prompt. Someone writes a few paragraphs describing the company, pastes in a handful of documents, and wires it to a chat box. It answers plausibly for the demo and then, at the first question that touches a real record, it invents one. The problem is not the model. The problem is that the assistant was told about the business rather than built from it.
Autoagentix Chat starts from the opposite premise. It is not a chatbot; it is the machine that produces them. It is early software, v0.1.0, in active development — so the honest framing is a design we are building out, not a finished product with outcomes we have measured. What the design already fixes is the gap between what an assistant claims and what it can trace.
Read the system four ways
The first move is not to write a prompt. It is to read the system — and to read it four ways rather than one. One source on its own is partial and often out of date; four sources cross-check each other, and the disagreements between them are usually where the real behaviour hides.
| Source | What it recovers |
|---|---|
| Repository (8 web frameworks) | What the application can do |
| Database (Postgres, MySQL) | What it actually holds |
| Live API | What it exposes |
| Documentation | What people believe about it |
Every fact the system records cites the file and line it came from. That single discipline is what separates a grounded assistant from a well-briefed one.
There is no floating claim that “the application manages invoices.” There is a fact, tied to the exact place in the codebase or document that justifies it.
Generate from facts, and refuse when you cannot
From the grounded picture, Autoagentix Chat generates the assistant itself: the tools it can call, the workflows it follows, the sub-agents it delegates to, and the permissions that bound all of it. Each is derived from a grounded fact, not a guess.
The more important half is what happens when there is no grounded fact. An operation whose effect cannot be determined is refused, with its reason stated, rather than attempted. Anything that mutates state stops for a person before it runs.
Most assistants are built to answer, so silence feels like failure. Autoagentix Chat is built so that a refusal with a reason is a correct output, and an unverifiable action is a gate, not a green light.
Prove it, then hand it over
An assistant that cannot be re-checked later is a liability the day the system it describes changes. So the proof ships inside the handover: a package-derived eval suite with a baseline, which the business can re-run after any change to be told exactly what regressed.
Ownership is the last step and the one that matters most. The handover exports the adapter, the corpus, the capability manifest, the eval report and the README. The business ends up holding its own files, not a hosted secret it rents. It is built on a plain stack — Python with FastAPI, HTMX, SSE, MCP and MongoDB — and exposes eight engine verbs to MCP clients such as Claude and ChatGPT, so an existing tool can drive it without a bespoke integration.
See how the four-way read, the citations and the handover fit together on the Autoagentix Chat product page, or book a call at cal.com/agentix-tech.
