IRC-A in production, part 2: a 46-minute agent, an AI assistant that broke the rules — and a $0.04 bill
It's 1:53 AM and I'm asking my CRM's chatbot: "what's the total amount we have parked in the Proposal stage?" Three seconds later it answers: $3,134.90. I follow up with just "and in prospecting?" — no context — and it breaks down the 10 open opportunities with amounts, close probability and estimated dates. A week earlier, this system was an n8n flow that only knew how to count contacts. This is…
It is 1:53 AM and the reporter asked their CRM's chatbot a question about the total amount in the Proposal stage. The chatbot responded in just three seconds with $3,134.90. This was after just 46 minutes, the reporter timed it, from the decision to build a new sales-reports agent and plug it into the live system, without any modifications to the existing components.
During this process, they discovered several bugs and vulnerabilities in the system, including a crossed URL in a .env file, a silent break in the FastMCP instance-to-handler binding, and a regression of a synchronous LLM call blocking Uvicorn's event loop bug from a previous version. The reporter emphasized that the true value of the experiment was in how quickly and cleanly they could add a new node, revealing weaknesses in the surrounding components.
They also shared a concerning incident where an AI coding assistant modified one of the agents, violating three design invariants, but the system still degraded safely with two informative rejections and no data corruption. This incident served as an involuntary penetration test, proving that the design-enforced token lockdown was effective.
Lastly, the reporter confessed to removing a previous fix they had implemented, as it introduced unnecessary coupling and hindered the system's ability to discover new tools.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
