Breadcrumbs

A guided builder concierge. From idea to build — one breadcrumb at a time.

The Problem

Inside most enterprises, there's a quiet bottleneck killing internal innovation: non-technical employees have ideas but no idea how to build them. The internal builder ecosystem is dense — 10–15 approved tools across no-code platforms, low-code app builders, AI agent platforms, IDE-based code-first tools, workflow automation, and BI/dashboarding. Each tool has its own access path, governance rules, skill floor, and data classification limits.

The result: employees give up on ideas, build in the wrong tool and hit a wall mid-project, burn weeks chasing unnecessary access, or build something non-compliant. Generic AI chatbots make it worse — they recommend tools outside the approved catalog, invent access paths, and skip governance entirely.

The Solution

A guided builder concierge that walks any employee — non-technical or experienced — step by step from "I have an idea" to "I have a working build," using only enterprise-approved tools, with governance baked in at every step. Named Breadcrumbs after the Hansel and Gretel fairy tale — because the agent's job is to leave a clear trail so users never get lost in the builder forest.

Unlike generic chatbots, Breadcrumbs answers four harder questions:

1. What's the right-sized build for my idea?
2. Which approved tool fits — and which don't?
3. What's going to block me, and how do I get unblocked?
4. How do I move forward safely (governance, data classification, permissions)?

Architecture

Key Architectural Decisions

What I Built

• A 4-section knowledge source (~6,000 words) covering the full enterprise builder ecosystem — approved tools, governance rules, deprecated tools to redirect from, and 4 artifact templates
• An 8-question structured intake that doubles as a governance check (data classification is the gate question — required before any recommendation)
• A hardened instructions block (~7,200 characters) enforcing pacing, grounding, length, tone, and edge cases — solved real drift issues like agent inventing tools and pulling unintended personal context
• A brand: name, tagline, visual identity (storybook illustration icon), and a fairy-tale-anchored narrative that gave the agent a personality leadership wanted to share
• A demo-ready agent that walks users from idea → recommendation → artifacts in roughly 30 minutes per session

Lessons Learned

• Grounding > intelligence. The agent doesn't need to be smart — it needs to refuse to be wrong. Strict source rules + an approved catalog beat any fancy reasoning.
• Instructions are the product. The knowledge source matters, but the instructions block is where the personality, pacing, and trust live.
• Stateless is a feature. Once I framed Breadcrumbs as a concierge (fresh every session) instead of a memory-based assistant, every governance question got easier.
• Brand is part of the build. Naming it Breadcrumbs (not "BuilderBot") gave the agent a story. That story is what made the demo land.
• Constraints make better products. A 10-hour deadline, an 8,000-character instructions cap, and a 12-tool catalog forced sharper decisions than a 6-week timeline would have.

Command Stack