HHACKGENCY
AI AGENTSAI DEVELOPMENT · 2026 · BUDGET MOINS DE 3 K€

The Agentic Tribe's internal AI intake-to-ticket engine

MEASURED RESULTS

-87%

Internal time to production

THE PROBLEM

Every custom ERP we ship for a client keeps generating feedback once it goes live: a bug, a missing field, a process that does not match how the team actually works. That feedback arrived everywhere except in a form an engineer could act on — WhatsApp messages, one-off emails, a client filling in a Google Form and a project lead re-typing it into a ticket by hand. Someone always had to read the raw complaint, figure out what was actually being asked, go back to the client for the missing details, and only then write a ticket precise enough for a developer, or one of our own coding agents, to pick up. That triage step sat squarely between "the client said something" and "an agent can build it," and it did not scale past one or two client ERPs without turning into a full-time job.

THE SOLUTION

We built TheAgent: a standalone intake service (Hono, TypeScript, Prisma, Postgres) that any client ERP embeds as a single "/demandes" page. The end user just describes their problem in plain language. TheAgent — a Claude-powered agent with adaptive thinking, asks one sharp question at a time, surfaces contradictions with how the client's own system actually works, and keeps digging until the request has a concrete problem, a desired outcome, and clear acceptance criteria. Only then does it file a structured ticket, automatically, into our internal ticketing system, where our engineering agents (backend, UI, DevOps) pick it up and open a pull request. Highlights - One conversational intake agent, reused identically across every client ERP — no per-client fork to maintain - Turns unstructured client feedback (chat, forms, one-liners) into a ticket with title, structured description, and request type, ready for an engineer or a coding agent - Streamed conversation (SSE) over a qualifying dialogue, not a static intake form — the agent keeps asking until the request is actually buildable - Zero-trust tenant boundary: client identity is bound server-side to the authenticated app, never derived from user input or model output — closes the cross-tenant ticket-injection risk by design (ADR 0002) - Short-lived JWT auth per client app, rotatable independently, never exposed to the browser - Straight pipeline from "client describes a problem" to a ticket already sitting in the same queue our own delivery agents work from

STACK AND TOOLS

The Agentic Tribe's internal AI intake-to-ticket engine: The Agentic Tribe project | Hackgency